refactor

- Add threading support implementation details
- Document ThreadPoolExecutor usage and thread safety
- Add model parameter implementation details
- Include testing results for both features

CMakeLists.txt

@@ -112,7 +112,6 @@ option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_
 option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)
 
 # 3rd party libs
-option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
@@ -197,9 +196,7 @@ add_subdirectory(src)
 
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
-    if (LLAMA_HTTPLIB)
-        add_subdirectory(vendor/cpp-httplib)
-    endif()
+    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)

build-xcframework.sh

@@ -449,10 +449,9 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
@@ -465,10 +464,9 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet

common/CMakeLists.txt

@@ -112,11 +112,7 @@ endif()
 
 # TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
 set(LLAMA_COMMON_EXTRA_LIBS build_info)
-
-if (LLAMA_HTTPLIB)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
-endif()
+set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)

common/download.cpp

@@ -19,9 +19,7 @@
 #include <thread>
 #include <vector>
 
-#if defined(LLAMA_USE_HTTPLIB)
 #include "http.h"
-#endif
 
 #ifndef __EMSCRIPTEN__
 #ifdef __linux__
@@ -142,8 +140,6 @@ std::pair<std::string, std::string> common_download_split_repo_tag(const std::st
     return {hf_repo, tag};
 }
 
-#if defined(LLAMA_USE_HTTPLIB)
-
 class ProgressBar {
     static inline std::mutex mutex;
     static inline std::map<const ProgressBar *, int> lines;
@@ -768,30 +764,6 @@ std::string common_docker_resolve_model(const std::string & docker) {
     }
 }
 
-#else
-
-common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-bool common_download_model(const common_params_model &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-std::string common_docker_resolve_model(const std::string &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-int common_download_file_single(const std::string &,
-                                const std::string &,
-                                const std::string &,
-                                bool,
-                                const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-#endif // defined(LLAMA_USE_HTTPLIB)
-
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();

convert_hf_to_gguf.py

@@ -2726,8 +2726,6 @@ class AfmoeModel(LlamaModel):
         super().set_gguf_parameters()
 
         # MoE parameters
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (n_shared_experts := self.hparams.get("num_shared_experts")) is not None:
             self.gguf_writer.add_expert_shared_count(n_shared_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
@@ -2749,7 +2747,7 @@ class AfmoeModel(LlamaModel):
         # Handle expert weights - they're already merged in the HF format
         # process the experts separately
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -4074,6 +4072,87 @@ class InternVisionModel(MmprojModel):
                 yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register(
+    "NemotronH_Nano_VL_V2",
+    "RADIOModel",
+)
+class NemotronNanoV2VLModel(MmprojModel):
+    # ViT-Huge architecture parameters for RADIO v2.5-h
+    _vit_hidden_size = 1280
+    _vit_intermediate_size = 5120
+    _vit_num_layers = 32
+    _vit_num_heads = 16
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        # RADIO config doesn't have standard ViT parameters, so they need to be constructed manually
+        vision_config = self.global_config.get("vision_config")
+        if vision_config is None:
+            return None
+        # Add ViT-H parameters
+        vision_config = {
+            **vision_config,
+            "hidden_size": self._vit_hidden_size,
+            "intermediate_size": self._vit_intermediate_size,
+            "num_hidden_layers": self._vit_num_layers,
+            "num_attention_heads": self._vit_num_heads,
+            "image_size": self.global_config.get("force_image_size", 512),
+        }
+        return vision_config
+
+    def set_gguf_parameters(self):
+        if "image_mean" not in self.preprocessor_config:
+            self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
+        if "image_std" not in self.preprocessor_config:
+            self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225]
+
+        super().set_gguf_parameters()
+        hparams = self.global_config
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
+        self.gguf_writer.add_vision_use_gelu(True)
+        downsample_ratio = hparams.get("downsample_ratio", 0.5)
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ".position_embd." in new_name or "pos_embed" in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "input_conditioner" in name:
+            return
+
+        # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
+        if "patch_generator.pos_embed" in name:
+            if not name.endswith(".weight"):
+                name += ".weight"
+            # Downsample position embeddings for fixed 512x512 image size
+            import torch.nn.functional as F
+            n_embd = self.hparams["hidden_size"]
+            image_size = self.global_config.get("force_image_size", 512)
+            patch_size = self.hparams["patch_size"]
+            target_patches_per_side = image_size // patch_size  # 32
+            max_patches_per_side = int((data_torch.shape[1]) ** 0.5)  # 128
+            if target_patches_per_side != max_patches_per_side:
+                # Reshape to grid, interpolate, flatten back
+                data_torch = data_torch.reshape(1, max_patches_per_side, max_patches_per_side, n_embd)
+                data_torch = data_torch.permute(0, 3, 1, 2).float()  # [1, n_embd, 128, 128]
+                data_torch = F.interpolate(data_torch, size=(target_patches_per_side, target_patches_per_side),
+                                           mode='bilinear', align_corners=True)
+                data_torch = data_torch.permute(0, 2, 3, 1)  # [1, 32, 32, n_embd]
+                data_torch = data_torch.reshape(1, target_patches_per_side * target_patches_per_side, n_embd)
+
+        # Reshape linear patch embedding to conv2d format for ggml_conv_2d
+        # From [n_embd, patch_size*patch_size*3] to [n_embd, 3, patch_size, patch_size]
+        if "patch_generator.embedder" in name:
+            patch_size = self.hparams["patch_size"]
+            n_embd = self.hparams["hidden_size"]
+            data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)
+
+        if name.startswith("vision_model.radio_model.model.") or name.startswith("mlp1."):
+            yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -4116,8 +4195,6 @@ class Qwen2MoeModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
             logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
@@ -4162,7 +4239,7 @@ class Qwen2MoeModel(TextModel):
             return
 
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -4913,13 +4990,13 @@ class PhiMoeModel(Phi3MiniModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
-        self.gguf_writer.add_expert_count(self.hparams["num_local_experts"])
+        self.gguf_writer.add_expert_used_count(self.find_hparam(["num_experts_per_tok", "num_experts_per_token"]))
+        self.gguf_writer.add_expert_count(self.find_hparam(["num_local_experts", "num_experts"]))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("block_sparse_moe.experts") != -1:
-            n_experts = self.hparams["num_local_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -5331,7 +5408,7 @@ class KimiLinearModel(TextModel):
 
         # process the experts separately
         if name.find("block_sparse_moe.experts") != -1:
-            n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -5926,12 +6003,13 @@ class NomicBertModel(BertModel):
         if "mlp.experts.bias" in name:
             return # Explicitly return.
 
+        n_experts = self.find_hparam(["num_local_experts", "num_experts"])
         if "mlp.experts.mlp.w1" in name:
-            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.view(n_experts, self.hparams["n_inner"], self.hparams["n_embd"])
             name += ".weight"
 
         if "mlp.experts.mlp.w2" in name:
-            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.view(n_experts, self.hparams["n_inner"], self.hparams["n_embd"])
             data_torch = data_torch.transpose(1, 2)
             name += ".weight"
 
@@ -5941,7 +6019,6 @@ class NomicBertModel(BertModel):
         super().set_gguf_parameters()
         if self.is_moe:
             self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"])
-            self.gguf_writer.add_expert_count(self.hparams["num_experts"])
             self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
 
     def _is_tokenizer_xlmroberta(self) -> bool:
@@ -7055,6 +7132,8 @@ class Mamba2Model(TextModel):
         if hparams is None:
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 hparams = json.load(f)
+        if "llm_config" in hparams:
+            hparams["text_config"] = hparams["llm_config"]
         super().__init__(dir_model, *args, hparams=hparams, **kwargs)
         self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
         self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
@@ -7176,8 +7255,8 @@ class JambaModel(TextModel):
         self.gguf_writer.add_ssm_state_size(d_state)
         self.gguf_writer.add_ssm_time_step_rank(dt_rank)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
-        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
+        self.gguf_writer.add_expert_count(self.find_hparam(["num_local_experts", "num_experts"]))
+        self.gguf_writer.add_expert_used_count(self.find_hparam(["num_experts_per_tok", "num_experts_per_token"]))
         self.gguf_writer.add_file_type(self.ftype)
 
     _experts: list[dict[str, Tensor]] | None = None
@@ -7195,7 +7274,7 @@ class JambaModel(TextModel):
 
         # process the experts separately
         if ".feed_forward.experts." in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
 
             assert bid is not None
 
@@ -7343,8 +7422,6 @@ class OlmoeModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_layer_norm_rms_eps(1e-5)
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -7352,7 +7429,7 @@ class OlmoeModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -7933,10 +8010,6 @@ class MiniMaxM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.MINIMAXM2
     _experts_cache: dict[int, dict[str, Tensor]] = {}
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.hparams["num_experts"] = self.hparams["num_local_experts"]
-
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
@@ -7949,7 +8022,7 @@ class MiniMaxM2Model(TextModel):
 
         # merge expert weights
         if 'experts' in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             expert_cache = self._experts_cache.setdefault(bid, {})
@@ -9154,7 +9227,6 @@ class ExaoneMoEModel(Exaone4Model):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
         moe_intermediate_size = self.hparams["moe_intermediate_size"]
         num_shared_experts = self.hparams["num_shared_experts"]
         self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
@@ -9195,7 +9267,7 @@ class ExaoneMoEModel(Exaone4Model):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
 
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9346,7 +9418,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         # case, the model architecture needs to be updated to a standard
         # "granite" or "granitemoe" model
         if not self._ssm_layers:
-            has_experts = self.find_hparam(["num_experts_per_tok"], optional=True)
+            has_experts = self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)
             new_arch = (
                 gguf.MODEL_ARCH.GRANITE_MOE
                 if has_experts else
@@ -9542,6 +9614,14 @@ class NemotronHModel(GraniteHybridModel):
             self.gguf_writer.add_add_bos_token(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision model and projector tensors for VLM models (handled by mmproj) (e.g., Nemotron Nano 12B v2 VL)
+        if name.startswith(("vision_model.", "mlp1.")):
+            return
+
+        # Strip language_model. prefix for VLM models (e.g., Nemotron Nano 12B v2 VL)
+        if name.startswith("language_model."):
+            name = name[len("language_model."):]
+
         if self.is_moe and bid is not None:
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -9636,7 +9716,6 @@ class BailingMoeModel(TextModel):
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_weights_scale(1.0)
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
@@ -9670,7 +9749,7 @@ class BailingMoeModel(TextModel):
             yield from super().modify_tensors(v,self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)
             return
         elif name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9741,7 +9820,6 @@ class BailingMoeV2Model(TextModel):
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_shared_feed_forward_length(hparams.get("moe_shared_expert_intermediate_size", hparams["moe_intermediate_size"] * hparams["num_shared_experts"]))
         self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
         self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
 
@@ -9752,7 +9830,7 @@ class BailingMoeV2Model(TextModel):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if "mlp.experts" in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -9798,8 +9876,6 @@ class GroveMoeModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
             logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
@@ -9820,7 +9896,7 @@ class GroveMoeModel(TextModel):
 
         # process the experts separately
         if name.find("chunk_experts") != -1:
-            n_experts = self.hparams["num_experts"] // 2 # see add_experts_per_group
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"]) // 2 # see add_experts_per_group
             assert bid is not None
 
             if self._chunk_experts is None:
@@ -9847,7 +9923,7 @@ class GroveMoeModel(TextModel):
             else:
                 return
         elif name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10240,7 +10316,6 @@ class HunYuanMoEModel(TextModel):
         super().set_gguf_parameters()
         hparams = self.hparams
 
-        self.gguf_writer.add_expert_count(hparams["num_experts"])
         self.gguf_writer.add_expert_shared_feed_forward_length(hparams["intermediate_size"])
 
         moe_intermediate_size = hparams["moe_intermediate_size"]
@@ -10283,7 +10358,7 @@ class HunYuanMoEModel(TextModel):
                 return
 
         if name.find("mlp.experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10325,16 +10400,9 @@ class LLaDAMoEModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
-
         if (expert_intermediate_size := self.hparams.get("expert_intermediate_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
 
-        # number of experts used per token (top-k)
-        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
-            self.gguf_writer.add_expert_used_count(n_experts_used)
-
         self.gguf_writer.add_mask_token_id(156895)
         self.gguf_writer.add_causal_attention(False)
         self.gguf_writer.add_diffusion_shift_logits(False)
@@ -10345,7 +10413,7 @@ class LLaDAMoEModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:
@@ -10682,7 +10750,6 @@ class LFM2MoeModel(TextModel):
 
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
         self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
         self.gguf_writer.add_leading_dense_block_count(self.hparams["num_dense_layers"])
         self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
@@ -10703,7 +10770,7 @@ class LFM2MoeModel(TextModel):
 
         # merge expert weights
         if 'experts' in name:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             expert_cache = self._experts_cache.setdefault(bid, {})
@@ -10813,9 +10880,9 @@ class SmallThinkerModel(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))) is not None:
+        if (n_experts := self.hparams.get("moe_num_primary_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
-        if (n_experts_used := self.hparams.get("num_experts_per_tok", self.hparams.get("moe_num_active_primary_experts"))) is not None:
+        if (n_experts_used := self.hparams.get("moe_num_active_primary_experts")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
         if (moe_intermediate_size := self.hparams.get("moe_ffn_hidden_size")) is not None:
             self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
@@ -10840,7 +10907,7 @@ class SmallThinkerModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         if name.find("experts") != -1:
-            n_experts = self.hparams.get("num_experts", self.hparams.get("moe_num_primary_experts"))
+            n_experts = self.hparams.get("moe_num_primary_experts") or self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:

docs/build-s390x.md

@@ -242,10 +242,10 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 |------------|-------------|------|-------|
 | FP32       | ✅           | ✅    | ❓     |
 | FP16       | ✅           | ✅    | ❓     |
-| BF16       | 🚫           | ✅    | ❓     |
+| BF16       | ✅           | ✅    | ❓     |
 | Q4_0       | ✅           | ❓    | ❓     |
 | Q4_1       | ✅           | ❓    | ❓     |
-| MXFP4      | 🚫           | ❓    | ❓     |
+| MXFP4      | ✅           | ❓    | ❓     |
 | Q5_0       | ✅           | ❓    | ❓     |
 | Q5_1       | ✅           | ❓    | ❓     |
 | Q8_0       | ✅           | ❓    | ❓     |
@@ -272,4 +272,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Feb 15, 2026.

examples/llama-eval/AGENTS.md

@@ -0,0 +1,190 @@
+# llama-eval Codebase Guidelines
+
+## Overview
+
+This directory contains Python evaluation tools for llama.cpp:
+- `llama-eval.py` - Main evaluation tool with multiple datasets (AIME, AIME2025, GSM8K, GPQA)
+- `llama-server-simulator.py` - Flask-based server simulator for testing
+- `test-simulator.sh` - Test script for the simulator
+
+## Build/Run Commands
+
+### Virtual Environment
+The project uses a virtual environment located at `venv/`:
+```bash
+source venv/bin/activate
+```
+
+### Running the Main Evaluator
+```bash
+python llama-eval.py \
+  --server http://127.0.0.1:8013 \
+  --model gpt-oss-20b-hf-low \
+  --dataset aime \
+  --n_cases 10 \
+  --grader-type llm \
+  --seed 42
+```
+
+### Running the Simulator (for testing)
+```bash
+python llama-server-simulator.py --port 8033 --success-rate 0.8
+```
+
+### Running Tests
+```bash
+./test-simulator.sh
+```
+
+## Code Style Guidelines
+
+### Imports
+- Standard library imports first (argparse, json, os, re, subprocess, sys, time)
+- Third-party imports (requests, tqdm, datasets, flask) after standard library
+- Relative imports not used
+- Group imports by category with blank line between groups
+
+### Formatting
+- 4-space indentation
+- Max line length: 125 characters (per parent project's .flake8)
+- Use double quotes for strings
+- Use triple double quotes for docstrings
+- Binary operators at the beginning of continued lines
+
+### Naming Conventions
+- Classes: PascalCase (e.g., `AimeDataset`, `Grader`, `Processor`)
+- Functions: snake_case (e.g., `normalize_number`, `get_prompt`)
+- Variables: snake_case (e.g., `question_text`, `correct_count`)
+- Constants: UPPER_SNAKE_CASE (e.g., `GRADER_PATTERNS`, `TEMPLATE_REGISTRY`)
+- Private methods: prefix with underscore (e.g., `_load_dataset`, `_grade_regex`)
+
+### Types
+- Use type hints for all function signatures
+- Import from `typing` module: `Dict`, `List`, `Optional`, `Any`, `Tuple`
+- Use `@dataclass` for data structures
+- Prefer `Optional[T]` over `Union[T, None]`
+
+### Error Handling
+- Use try/except for network requests and file operations
+- Return `None` or `False` on errors when appropriate
+- Use `ValueError` for invalid arguments
+- Use `FileNotFoundError` for missing files
+- CLI scripts should handle exceptions gracefully
+
+### Dataclasses
+- Use `@dataclass` for structured data
+- Define fields with explicit types
+- Use `Optional[T]` for nullable fields
+- Provide default values where appropriate
+
+### String Formatting
+- Use f-strings for formatting (Python 3.6+)
+- Use triple double quotes for multi-line strings
+- Escape backslashes in regex patterns: `r'\\boxed{(\d+)}'`
+
+### File Paths
+- Use `pathlib.Path` instead of string paths
+- Create directories with `mkdir(parents=True, exist_ok=True)`
+- Use `Path.home()` for user home directory
+
+### Logging
+- Use `print()` for user-facing output
+- Use `sys.stderr` for debug logging
+- Simulator writes debug logs to `/tmp/simulator-debug.log`
+
+### Testing
+
+- Test script uses bash with `set -e` for strict error handling
+- Simulator runs in background with PID tracking
+- Tests verify correct answers, error cases, and edge cases
+- Use `curl` for HTTP testing in shell scripts
+
+### Whitespace Cleanup
+- Remove trailing whitespace from all lines
+- When making edits, do not leave trailing whitespace
+
+## Dataset Support
+
+### AIME Dataset
+- 90 questions from 2025 AIME competition
+- Answers in `\boxed{answer}` format
+- Supports regex, CLI, and LLM grading
+
+### AIME2025 Dataset
+- 30 questions from 2025 AIME I & II
+- Answers in `\boxed{answer}` format
+- Requires loading two config parts
+
+### GSM8K Dataset
+- 7473 math word problems
+- Answers numeric values with `####` separator
+- Supports regex, CLI, and LLM grading
+
+### GPQA Dataset
+- 198 questions from GPQA Diamond
+- Multiple choice with shuffled options (A, B, C, D)
+- **Requires LLM grader** (returns letter A/B/C/D)
+
+## Grading Types
+
+### Regex Grader
+- Built-in patterns per dataset
+- Prioritizes `\boxed{}` for AIME datasets
+- Extracts last number for GSM8K
+
+### CLI Grader
+- External script interface
+- Call: `grader.sh --answer <pred> --expected <gold>`
+- Exit code 0 = correct, non-zero = incorrect
+
+### LLM Grader
+- Uses judge model for answer extraction
+- Includes few-shot examples
+- Case-insensitive comparison
+- Required for GPQA
+
+## Configuration
+
+### Sampling Parameters (Optional)
+- `--temperature`: Sampling temperature
+- `--top-k`: Top K sampling
+- `--top-p`: Top P sampling
+- `--min-p`: Min P sampling
+- Only passed to API if explicitly specified
+
+### Default Values
+- `--n_predict`: -1 (infinite)
+- `--grader-type`: llm
+- `--seed`: 1234
+- `--threads`: 32
+- `--output`: llama-eval-state.json
+
+## Output Format
+
+### Progress Table
+- Shows task ID, dataset, prompt (truncated to 43 chars), expected answer, status
+- Uses `tqdm` for progress bars
+
+### Results Summary
+- Format: `Results: X/Y correct (Z%)`
+- Displayed after all tasks complete
+
+### JSON Output
+- Complete eval state saved to output file
+- Contains: task IDs, correctness, prompts, extracted answers, sampling config
+- Uses `dataclasses.asdict()` for serialization
+
+## HuggingFace Datasets
+
+- Cache directory: `~/.cache/huggingface/datasets`
+- Set via `HF_DATASETS_CACHE` environment variable
+- Telemetry disabled via `HF_HUB_DISABLE_TELEMETRY=1`
+- Datasets loaded with `datasets.load_dataset()`
+
+## Flask Simulator
+
+- Runs on configurable port (default: 5000)
+- Endpoint: `/v1/chat/completions` (OpenAI-compatible)
+- Uses Dice coefficient for question matching
+- Configurable success rate for testing
+- Debug logs to `/tmp/simulator-debug.log`

examples/llama-eval/IMPLEMENTATION.md

@@ -0,0 +1,94 @@
+# llama-eval Implementation Summary
+
+## Overview
+
+Simple evaluation tool for llama.cpp with support for multiple datasets (AIME, GSM8K, GPQA) and flexible grading (regex, CLI, LLM).
+
+## Key Features
+
+- **Multiple Datasets**: AIME, GSM8K, GPQA with proper answer extraction
+- **Flexible Grading**: Regex, CLI, or LLM-based grading
+- **Parallel Processing**: Configurable thread count for concurrent requests
+- **Sampling Parameters**: Temperature, Top K, Top P, Min P (optional)
+- **Real-time Feedback**: Progress tracking with detailed output
+- **JSON Output**: Complete eval state saved for debugging
+- **GPQA Support**: Answer shuffling with reproducible results
+
+## Architecture
+
+### Eval State
+```python
+@dataclass
+class EvalState:
+    id: str
+    tasks: List[str]
+    task_states: Dict[str, Dict[str, Any]]
+    sampling_config: Dict[str, Any]
+```
+
+### Processor
+- Handles processing, grading, and state management
+- Thread-safe concurrent execution
+- Configurable sampling parameters
+
+### Grader
+- Abstract grading interface supporting multiple types
+- Regex grader with dataset-specific patterns
+- CLI grader with external script interface
+- LLM grader with configurable server and model
+
+### Datasets
+- `AimeDataset`: 90 AIME 2025 questions
+- `Aime2025Dataset`: 30 AIME 2025 I & II questions
+- `Gsm8kDataset`: 7473 math word problems
+- `GpqaDataset`: 198 GPQA Diamond questions with shuffling
+
+## Configuration
+
+### Sampling Parameters (Optional)
+- `--temperature`: Sampling temperature
+- `--top-k`: Top K sampling
+- `--top-p`: Top P sampling
+- `--min-p`: Min P sampling
+- Only passed if explicitly specified
+
+### Grading Types
+- **regex**: Built-in patterns for each dataset
+- **cli**: External script with `--answer` and `--expected` args
+- **llm**: LLM-based extraction with few-shot examples and configurable server/model
+
+### Dataset Requirements
+- **AIME**: Supports regex, CLI, or LLM grader
+- **AIME2025**: Supports regex, CLI, or LLM grader
+- **GSM8K**: Supports regex, CLI, or LLM grader
+- **GPQA**: Requires LLM grader
+
+## Output Format
+
+### Progress Table
+```
+  Task ID             Dataset  Prompt (first 43 chars)                        Expected    Status
+  aime_000_001         AIME   Complete the following reactions and sel...    A          pending
+```
+
+### Results Summary
+```
+============================================================
+Results: 8/10 correct (80.0%)
+============================================================
+```
+
+### JSON Output
+Complete eval state with task IDs, correctness, prompts, extracted answers, and sampling configuration.
+
+## Technical Details
+
+- Default max tokens: -1 (infinite)
+- Default grader type: llm
+- Default seed: 1234
+- Default threads: 32
+- Prompt truncation: First 43 chars + padding + "..."
+- Response truncation: Last 10 lines for grading
+- GPQA requires LLM grader (returns letter A/B/C/D)
+- Judge model defaults to evaluated model if not specified
+- Sample answers defined in SAMPLE_ANSWERS dict for few-shot learning

examples/llama-eval/README.md

@@ -0,0 +1,112 @@
+# llama-eval Evaluation Tool
+
+Simple evaluation tool for llama.cpp with support for multiple datasets.
+
+## Features
+
+- **Multiple Datasets**: AIME, GSM8K, GPQA
+- **Flexible Grading**: Regex, CLI, or LLM-based grading
+- **Parallel Processing**: Configurable thread count
+- **Real-time Feedback**: Progress tracking with detailed output
+- **Sampling Parameters**: Temperature, Top K, Top P, Min P
+- **JSON Output**: Complete eval state saved for debugging
+
+## Usage
+
+```bash
+python llama-eval.py \
+  --server http://127.0.0.1:8013 \
+  --model gpt-oss-20b-hf-low \
+  --judge-model gpt-oss-20b-hf-medium \
+  --dataset aime \
+  --n_cases 10 \
+  --grader-type llm \
+  --seed 42
+```
+
+## CLI Arguments
+
+- `--server`: llama-server URL (default: http://127.0.0.1:8013)
+- `--model`: Model name for evaluation (default: llama)
+- `--judge-model`: Model name for LLM judge (default: same as main model)
+- `--judge-server`: Server URL for LLM judge (default: same as main server)
+- `--dataset`: Dataset type (aime, aime2025, gsm8k, gpqa)
+- `--n_cases`: Number of cases to evaluate (default: all)
+- `--n_predict`: Max tokens to predict per prompt (default: -1, infinite)
+- `--temperature`: Sampling temperature (default: not passed)
+- `--top-k`: Top K sampling (default: not passed)
+- `--top-p`: Top P sampling (default: not passed)
+- `--min-p`: Min P sampling (default: not passed)
+- `--threads`: Number of threads for parallel requests (default: 32)
+- `--verbose`: Show detailed output for each case
+- `--output`: Output file for eval state (default: llama-eval-state.json)
+- `--grader-type`: Grader type (regex, cli, llm, default: llm)
+- `--grader-script`: Path to CLI grader script (required for --grader-type cli)
+- `--seed`: Random seed for shuffling (default: 1234)
+
+## Datasets
+
+### AIME
+- 90 questions from 2025 AIME competition
+- Answers in boxed format: `\boxed{answer}`
+- Requires regex grader or LLM grader
+
+### AIME2025
+- 30 questions from 2025 AIME I & II competitions
+- Answers in boxed format: `\boxed{answer}`
+- Supports regex, CLI, or LLM grader
+
+### GSM8K
+- 7473 math word problems
+- Answers are numeric values
+- Requires regex grader or LLM grader
+
+### GPQA
+- 198 questions from GPQA Diamond dataset
+- Multiple choice with shuffled options
+- Requires LLM grader (returns letter A, B, C, or D)
+
+## Grading Types
+
+### Regex Grader
+Built-in patterns for different datasets:
+- AIME: `\boxed{(\d+)}|\b(\d+)\b`
+- AIME2025: `\boxed{(\d+)}|\b(\d+)\b`
+- GSM8K: `\b(\d+)\b`
+- GPQA: Letter extraction (A, B, C, D)
+
+### CLI Grader
+External script interface:
+```bash
+./grader.sh --answer <pred> --expected <gold>
+```
+Returns exit code 0 if correct, non-zero if incorrect.
+
+### LLM Grader
+Uses LLM to extract and compare answers:
+- Configurable server and model
+- Includes few-shot examples from sample answers
+- Case-insensitive comparison
+- Required for GPQA dataset
+
+## Output
+
+### Progress Table
+```
+  Task ID             Dataset  Prompt (first 43 chars)                        Expected    Status
+  aime_000_001         AIME   Complete the following reactions and sel...    A          pending
+```
+
+### Results
+```
+============================================================
+Results: 8/10 correct (80.0%)
+============================================================
+```
+
+### JSON Output
+Complete eval state saved to output file with:
+- Task IDs and correctness status
+- Prompts and extracted answers
+- Sampling configuration
+- Processing metadata

examples/llama-eval/llama-server-simulator-README.md

@@ -0,0 +1,36 @@
+# llama-server-simulator
+
+Standalone Python script simulating llama-server HTTP endpoint for testing.
+
+## Features
+
+- HTTP Server with OpenAI-compatible `/v1/chat/completions` endpoint
+- AIME Dataset Integration - Loads 90 questions from HuggingFace
+- Intelligent Question Matching - Uses exact matching, LaTeX removal, and Levenshtein distance
+- Configurable Success Rate - Control correct/wrong answer generation (0-1)
+- Debug Logging - Troubleshoot matching issues
+
+## Usage
+
+```bash
+python llama-server-simulator.py --success-rate 0.8
+```
+
+## Arguments
+
+- `--success-rate`: Probability of returning correct answer (0.0-1.0, default: 0.8)
+- `--port`: Server port (default: 8033)
+- `--debug`: Enable debug logging (default: False)
+
+## Testing
+
+```bash
+./test-simulator.sh
+```
+
+## Implementation Details
+
+- Uses Levenshtein distance for partial matching (threshold: 0.3)
+- Automatic caching via HuggingFace datasets library
+- Wrong answers generated by incrementing expected answer
+- Debug output written to stderr

examples/llama-eval/llama-server-simulator.py

@@ -0,0 +1,283 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import random
+import re
+import time
+import sys
+import os
+from typing import Dict, List, Optional
+from dataclasses import dataclass, asdict
+from pathlib import Path
+
+import datasets
+from flask import Flask, request, jsonify
+
+# Set cache directory for HuggingFace datasets
+cache_dir = Path.home() / ".cache" / "huggingface" / "datasets"
+cache_dir.mkdir(parents=True, exist_ok=True)
+os.environ["HF_DATASETS_CACHE"] = str(cache_dir)
+
+def dice(s1: str, s2: str) -> float:
+    """Calculate Dice coefficient between two strings based on bigram overlap."""
+    if not s1 and not s2:
+        return 1.0
+
+    def _bigrams(s: str):
+        return [s[i : i + 2] for i in range(len(s) - 1)]
+
+    bigrams1 = _bigrams(s1)
+    bigrams2 = _bigrams(s2)
+
+    if not bigrams1 and not bigrams2:
+        return 1.0
+
+    from collections import Counter
+
+    freq1 = Counter(bigrams1)
+    freq2 = Counter(bigrams2)
+
+    intersection = sum(min(freq1[bg], freq2[bg]) for bg in freq1)
+    dice_coeff = 2 * intersection / (len(bigrams1) + len(bigrams2))
+    return dice_coeff
+
+def debug_log(message: str):
+    """Log debug messages to both stdout and a file"""
+    print(message, file=sys.stderr)
+    with open("/tmp/simulator-debug.log", "a") as f:
+        f.write(message + "\n")
+
+app = Flask(__name__)
+
+@dataclass
+class EvalState:
+    id: str
+    tasks: List[str]
+    task_states: Dict[str, Dict]
+    sampling_config: Dict
+
+def normalize_number(s: str) -> Optional[int]:
+    match = re.match(r"\d+", s)  # match digits from the start
+    if not match:
+        return None
+    return int(match.group(0))
+
+class AimeDataset:
+    def __init__(self, split: str = "train"):
+        self.split = split
+        self.questions: List[Dict] = []
+        self._load_dataset()
+
+    def _load_dataset(self):
+        print(f"Loading AIME dataset (split: {self.split})...")
+
+        cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "AI-MO___aimo-validation-aime" / "default" / "0.0.0"
+        if cache_path.exists():
+            print(f"Using cached dataset from {cache_path}")
+            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split, cache_dir=str(cache_path))
+        else:
+            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split)
+
+        self.questions = list(ds)
+        print(f"AIME dataset loaded: {len(self.questions)} questions")
+
+    def find_question(self, request_text: str) -> Optional[Dict]:
+        best_match = None
+        best_distance = -1
+        best_index = -1
+
+        for i, question in enumerate(self.questions):
+            question_text = question["problem"]
+            request_lower = request_text.lower()
+            question_lower = question_text.lower()
+
+            # Exact match
+            if question_lower == request_lower:
+                debug_log(f"DEBUG: Found exact match at index {i}")
+                return question
+
+            # Remove LaTeX formatting for more flexible matching
+            question_no_latex = re.sub(r'\$[^$]+\$', '', question_text)
+            if question_no_latex.lower() == request_lower:
+                debug_log(f"DEBUG: Found match (no LaTeX) at index {i}")
+                return question
+
+            # Calculate Levenshtein distance for partial matches
+            # Only consider if request is at least 50% of question length
+            if len(request_lower) >= len(question_lower) * 0.5:
+                distance = dice(question_lower, request_lower)
+
+                if distance > best_distance:
+                    best_distance = distance
+                    best_match = question
+                    best_index = i
+
+        if best_match and best_distance > 0.3:  # Threshold for partial match
+            debug_log(f"DEBUG: Found best partial match at index {best_index} with distance {best_distance:.3f}")
+            return best_match
+
+        debug_log(f"DEBUG: No matching question found for: {request_text[:100]}...")
+        return None
+
+    def get_answer(self, question: Dict) -> str:
+        answer = question["answer"]
+        if isinstance(answer, str):
+            normalized = normalize_number(answer)
+            return str(normalized) if normalized is not None else answer
+        return str(answer)
+
+class Simulator:
+    def __init__(
+        self,
+        port: int = 8033,
+        host: str = "localhost",
+        success_rate: float = 0.8,
+        dataset_split: str = "train"
+    ):
+        self.port = port
+        self.host = host
+        self.success_rate = success_rate
+        self.dataset = AimeDataset(dataset_split)
+        self.eval_state = EvalState(
+            id="aime-2025",
+            tasks=["aime"],
+            task_states={},
+            sampling_config={"temperature": 0, "max_tokens": 2048}
+        )
+
+    def _generate_response(
+        self,
+        question: Dict,
+        should_be_correct: bool
+    ) -> Dict:
+        expected_answer = self.dataset.get_answer(question)
+
+        if should_be_correct:
+            response_text = expected_answer
+        else:
+            response_text = self._generate_wrong_answer(question)
+
+        return {
+            "id": f"chatcmpl-{int(time.time())}",
+            "object": "chat.completion",
+            "created": int(time.time()),
+            "model": "llama",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": response_text
+                    },
+                    "finish_reason": "stop"
+                }
+            ],
+            "usage": {
+                "prompt_tokens": 100,
+                "completion_tokens": 50,
+                "total_tokens": 150
+            }
+        }
+
+    def _generate_wrong_answer(self, question: Dict) -> str:
+        expected_answer = self.dataset.get_answer(question)
+
+        if expected_answer.isdigit():
+            wrong_answer = str(int(expected_answer) + 1)
+        else:
+            wrong_answer = expected_answer + " (wrong)"
+
+        return wrong_answer
+
+    def _process_request(self, request_data: Dict) -> Dict:
+        messages = request_data.get("messages", [])
+        if not messages:
+            return {"error": "No messages in request"}
+
+        request_text = messages[0].get("content", "")
+        debug_log(f"DEBUG: Received request with content: {request_text[:150]}...")
+
+        question = self.dataset.find_question(request_text)
+        if not question:
+            debug_log(f"DEBUG: find_question returned None")
+            return {"error": "No matching question found"}
+
+        should_be_correct = random.random() < self.success_rate
+
+        response = self._generate_response(question, should_be_correct)
+
+        task_id = "aime"
+        self.eval_state.task_states[task_id] = {
+            "correct": should_be_correct,
+            "expected": self.dataset.get_answer(question),
+            "predicted": response["choices"][0]["message"]["content"]
+        }
+
+        return response
+
+@app.route('/v1/chat/completions', methods=['POST'])
+def chat_completions():
+    try:
+        request_data = request.get_json()
+
+        if not request_data:
+            return jsonify({"error": "Invalid JSON"}), 400
+
+        response = simulator._process_request(request_data)
+
+        return jsonify(response)
+
+    except Exception as e:
+        print(f"Error processing request: {e}")
+        return jsonify({"error": str(e)}), 500
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="llama-server simulator for testing eval scripts"
+    )
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=8033,
+        help="Server port (default: 8033)"
+    )
+    parser.add_argument(
+        "--host",
+        type=str,
+        default="localhost",
+        help="Server host (default: localhost)"
+    )
+    parser.add_argument(
+        "--success-rate",
+        type=float,
+        default=0.8,
+        help="Success rate 0-1 (default: 0.8)"
+    )
+    parser.add_argument(
+        "--dataset-split",
+        type=str,
+        default="train",
+        help="AIME dataset split to use (default: train)"
+    )
+
+    args = parser.parse_args()
+
+    global simulator
+    simulator = Simulator(
+        port=args.port,
+        host=args.host,
+        success_rate=args.success_rate,
+        dataset_split=args.dataset_split
+    )
+
+    print("\n=== llama-server-simulator ===")
+    print(f"Server running on http://{args.host}:{args.port}")
+    print(f"Success rate: {args.success_rate}")
+    print(f"AIME dataset loaded: {len(simulator.dataset.questions)} questions")
+    print("\nPress Ctrl+C to stop\n")
+
+    app.run(host=args.host, port=args.port, debug=False)
+
+if __name__ == "__main__":
+    main()

examples/llama-eval/test-simulator.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+
+set -e
+
+# Get the directory where this script is located
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+echo "=== llama-server-simulator Test Script ==="
+echo ""
+
+PORT=8033
+SUCCESS_RATE=0.8
+TEST_PORT=8034
+
+echo "Starting simulator on port $PORT with success rate $SUCCESS_RATE..."
+source "$SCRIPT_DIR/venv/bin/activate"
+python3 "$SCRIPT_DIR/llama-server-simulator.py" --port $PORT --success-rate $SUCCESS_RATE > /tmp/simulator-test.log 2>&1 &
+SIMULATOR_PID=$!
+
+echo "Waiting for simulator to start..."
+sleep 5
+
+# Helper function to make a request and extract the answer
+make_request() {
+  local question="$1"
+  curl -s -X POST http://localhost:$PORT/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d "{
+      \"model\": \"llama\",
+      \"messages\": [
+        {\"role\": \"user\", \"content\": \"$question\"}
+      ],
+      \"temperature\": 0,
+      \"max_tokens\": 2048
+    }" | python3 -c "import sys, json; data = json.load(sys.stdin); print(data.get('choices', [{}])[0].get('message', {}).get('content', data.get('error', 'No response')))"
+}
+
+# Test question (repeated in multiple tests)
+TEST_QUESTION="Quadratic polynomials P(x) and Q(x) have leading coefficients 2 and -2, respectively. The graphs of both polynomials pass through the two points (16,54) and (20,53). Find P(0) + Q(0)."
+
+echo ""
+echo "=== Test 1: Correct Answer ==="
+echo "Sending request with known question..."
+answer=$(make_request "$TEST_QUESTION")
+echo "Answer: $answer"
+echo "Expected: 116"
+echo "Correct: $([ "$answer" == "116" ] && echo "Yes" || echo "No")"
+
+echo ""
+echo "=== Test 2: Wrong Answer ==="
+echo "Sending request with known question (success rate 0.0)..."
+answer=$(make_request "$TEST_QUESTION")
+echo "Answer: $answer"
+echo "Expected: 116"
+echo "Correct: $([ "$answer" == "116" ] && echo "Yes" || echo "No")"
+
+echo ""
+echo "=== Test 3: No Matching Question ==="
+echo "Sending request with non-matching text..."
+response=$(make_request "What is the capital of France?")
+echo "Response: $response"
+echo "Expected: No matching question found"
+echo "Correct: $([ "$response" == "No matching question found" ] && echo "Yes" || echo "No")"
+
+echo ""
+echo "=== Test 4: Success Rate Verification ==="
+echo "Sending 10 requests to test success rate..."
+correct_count=0
+for i in {1..10}; do
+  answer=$(make_request "$TEST_QUESTION")
+  if [ "$answer" == "116" ]; then
+    correct_count=$((correct_count + 1))
+  fi
+  echo "  Request $i: Answer = $answer"
+done
+echo "Correct answers: $correct_count/10"
+echo "Expected: ~8/10 (80% success rate)"
+echo "Success rate: $(echo "scale=1; $correct_count * 10" | bc)%"
+
+echo ""
+echo "=== Test Complete ==="
+echo "Stopping simulator..."
+kill $SIMULATOR_PID 2>/dev/null
+wait $SIMULATOR_PID 2>/dev/null || true
+
+echo "Simulator stopped."

ggml/src/ggml-cpu/CMakeLists.txt

@@ -569,27 +569,24 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             cmake_policy(SET CMP0135 NEW)
         endif()
 
+        # TODO: Use FetchContent_MakeAvailable with EXCLUDE_FROM_ALL after bumping minimum CMake version to 3.28+
+        # Using FetchContent_Populate instead to avoid EXCLUDE_FROM_ALL which requires CMake 3.28
         FetchContent_Declare(KleidiAI_Download
             URL ${KLEIDIAI_DOWNLOAD_URL}
             DOWNLOAD_EXTRACT_TIMESTAMP NEW
             URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
 
-        FetchContent_MakeAvailable(KleidiAI_Download)
         FetchContent_GetProperties(KleidiAI_Download
             SOURCE_DIR  KLEIDIAI_SRC
             POPULATED   KLEIDIAI_POPULATED)
 
         if (NOT KLEIDIAI_POPULATED)
-            message(FATAL_ERROR "KleidiAI source downloaded failed.")
+            FetchContent_Populate(KleidiAI_Download)
+            FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
         endif()
 
         add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
 
-        # Remove kleidiai target after fetching it
-        if (TARGET kleidiai)
-            set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
-        endif()
-
         list(APPEND GGML_CPU_SOURCES
             ggml-cpu/kleidiai/kleidiai.cpp
             ggml-cpu/kleidiai/kernels.cpp

ggml/src/ggml-cpu/common.h

@@ -6,8 +6,8 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 
-#define GGML_FA_TILE_Q  32
-#define GGML_FA_TILE_KV 16
+#define GGML_FA_TILE_Q  64
+#define GGML_FA_TILE_KV 64
 
 #ifdef __cplusplus
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2874,8 +2874,8 @@ struct ggml_cplan ggml_graph_plan(
                         const int64_t DV = node->src[2]->ne[0];
 
                         // Tiled flash attention scratch (tile sizes defined in common.h)
-                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
-                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + K_f32 + padding
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV + GGML_FA_TILE_KV*DK)*n_tasks;
 
                         // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
                         // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
@@ -2947,7 +2947,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         /*.use_ref    =*/ cplan->use_ref,
     };
 
-    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
@@ -2974,7 +2978,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
-    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     ggml_barrier(state->threadpool);
 

ggml/src/ggml-cpu/ops.cpp

@@ -3,6 +3,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "binary-ops.h"
+#include "simd-gemm.h"
 #include "ggml.h"
 #include "unary-ops.h"
 #include "vec.h"
@@ -8389,10 +8390,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     GGML_ASSERT(k->type == v->type);
     const ggml_type kv_type = k->type;
 
-    const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type);
-    const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float;
-    const ggml_vec_dot_t    kv_vec_dot    = kv_type_traits_cpu->vec_dot;
-    const size_t kv_type_size = ggml_type_size(kv_type);
 
     // broadcast factors
     const int64_t rk2 = neq2/nek2;
@@ -8424,8 +8421,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
     static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
 
-    GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ");
-
     int ir = ir0;
     while (ir < ir1) {
         // q indices for the start of this tile
@@ -8452,18 +8447,20 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         }
 
         // Per-thread scratch layout:
-        // Q_q:    Q_TILE_SZ * DK (converted Q tile in KV type)
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile — F32 for GEMM, KV type for scalar)
         // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
         // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
         // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
-        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion)
-        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32);
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile)
+        // K_f32:  KV_TILE_SZ * DK (F32 buffer for K tile — GEMM path)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + KV_TILE_SZ*DK + CACHE_LINE_SIZE_F32);
 
         void  * Q_q    = base;
         float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
         float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
         float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
-        float * V32    = VKQ32 + Q_TILE_SZ * DV;  // F32 buffer for V tile
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;
+        float * K_f32  = V32 + KV_TILE_SZ * DV;
 
         memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
         memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
@@ -8476,28 +8473,38 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         const int iv3 = iq3 / rv3;
         const int iv2 = iq2 / rv2;
 
-        for (int tq = 0; tq < tile_rows; tq++) {
-            const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
-            kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK);
-        }
-        // Zero-pad remaining rows
-        for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
-            memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size);
+        {
+            float * Q_f32 = (float *)Q_q;
+            for (int tq = 0; tq < tile_rows; tq++) {
+                const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+                memcpy(Q_f32 + tq * DK, pq, DK * sizeof(float));
+            }
+            for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+                memset(Q_f32 + tq * DK, 0, DK * sizeof(float));
+            }
         }
 
+        memset(K_f32, 0, DK * KV_TILE_SZ * sizeof(float));
+        memset(V32,   0, KV_TILE_SZ * DV * sizeof(float));
+
         for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, nek1 - ic);
 
             // skip the tile entirely if all the masks are -inf
             if (mask) {
                 bool can_skip = true;
                 for (int tq = 0; tq < tile_rows; tq++) {
                     const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    for (int tk = 0; tk < kv_tile; tk++) {
                         mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
                         if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
                             can_skip = false;
                         }
                     }
+                    // Pad remaining mask entries with -inf
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
 
                 if (can_skip) {
@@ -8505,13 +8512,32 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 }
             }
 
-            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                const void * q_row = (const char *)Q_q + tq * DK * kv_type_size;
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3);
-                    float s;
-                    kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1);
-                    KQ[tq * KV_TILE_SZ + tk] = s * scale;
+            // Pack K tile transposed: K_f32[dk][kv] so KV_TILE is contiguous (SIMD dim)
+            // Zero-pad the last tile so the GEMM always operates on KV_TILE_SZ columns
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * k_data = (const char *)k->data + (ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
+                if (kv_type == GGML_TYPE_F16) {
+                    const ggml_fp16_t * k_f16 = (const ggml_fp16_t *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = GGML_CPU_FP16_TO_FP32(k_f16[dk]);
+                    }
+                } else {
+                    const float * k_f32_src = (const float *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = k_f32_src[dk];
+                    }
+                }
+            }
+            memset(KQ, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+            simd_gemm(KQ, (const float *)Q_q, K_f32, Q_TILE_SZ, DK, KV_TILE_SZ);
+            ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, scale);
+
+            // Set padded KQ entries to -inf so softmax gives them zero weight
+            if (kv_tile < KV_TILE_SZ) {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        KQ[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
             }
 
@@ -8551,33 +8577,22 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
             }
 
-            // Convert V tile to F32 first (if F16), then do MAD
-            // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster.
-            // TODO: on ARM, native f16 should be faster
-            if (kv_type == GGML_TYPE_F16) {
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                    ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV);
-                }
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p);
-                    }
-                }
-            } else {
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                        ggml_vec_mad_f32(DV, vkq_row, v_row, p);
-                    }
+            // V accumulation: VKQ32 += softmax(KQ) * V
+            // Pack V tile to contiguous F32, zero-padded
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
+                if (kv_type == GGML_TYPE_F16) {
+                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                } else {
+                    memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
                 }
             }
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                if (skip[tq]) {
+                    memset(KQ + tq * KV_TILE_SZ, 0, KV_TILE_SZ * sizeof(float));
+                }
+            }
+            simd_gemm(VKQ32, KQ, V32, Q_TILE_SZ, KV_TILE_SZ, DV);
         }
 
         // sinks (apply only to valid rows in the tile)
@@ -8794,15 +8809,15 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
         const int64_t dr = (nr + nchunk - 1) / nchunk;
 
-        static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
         static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
-        const bool use_tiled = !use_ref &&
+        bool use_tiled = !use_ref &&
                                (q->type == GGML_TYPE_F32 &&
                                 kv_is_f32_or_f16 &&
                                 k->type == v->type &&
-                                nek1 % KV_TILE_SZ == 0 &&
                                 neq1 >= Q_TILE_SZ);
-
+#ifdef GGML_SIMD
+        use_tiled &= (DV % GGML_F32_EPR == 0);
+#endif
         int current_chunk = ith;
 
         while (current_chunk < nchunk) {

ggml/src/ggml-cpu/simd-gemm.h

@@ -0,0 +1,136 @@
+#pragma once
+
+// Computes C[M x N] += A[M x K] * B[K x N]
+
+#include "simd-mappings.h"
+
+// TODO: add support for sizeless vector types
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+
+// TODO: untested on avx512
+// These are in units of GGML_F32_EPR
+#if defined(__AVX512F__) || defined (__ARM_NEON__)
+    static constexpr int GEMM_RM = 4;
+    static constexpr int GEMM_RN = 4; // 16+4+1 = 25/32
+#elif defined(__AVX2__) || defined(__AVX__)
+    static constexpr int GEMM_RM = 6;
+    static constexpr int GEMM_RN = 2; // 12+2+1 = 15/16
+#else
+    static constexpr int GEMM_RM = 2;
+    static constexpr int GEMM_RN = 2;
+#endif
+
+template <int RM, int RN>
+static inline void simd_gemm_ukernel(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    GGML_F32_VEC acc[RM][RN];
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            acc[i][r] = GGML_F32_VEC_LOAD(C + i * N + r * KN);
+        }
+    }
+
+    for (int64_t kk = 0; kk < K; kk++) {
+        GGML_F32_VEC Bv[RN];
+        for (int r = 0; r < RN; r++) {
+            Bv[r] = GGML_F32_VEC_LOAD(B + kk * N + r * KN);
+        }
+        for (int64_t i = 0; i < RM; i++) {
+            GGML_F32_VEC p = GGML_F32_VEC_SET1(A[i * K + kk]);
+            for (int r = 0; r < RN; r++) {
+                acc[i][r] = GGML_F32_VEC_FMA(acc[i][r], Bv[r], p);
+            }
+        }
+    }
+
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            GGML_F32_VEC_STORE(C + i * N + r * KN, acc[i][r]);
+        }
+    }
+}
+
+// C[M x N] += A[M x K] * B[K x N]
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    int64_t ii = 0;
+    for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<GEMM_RM, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<GEMM_RM, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            for (int64_t i = 0; i < GEMM_RM; i++) {
+                float a = C[i * N + jj];
+                for (int64_t kk = 0; kk < K; kk++) {
+                    a += A[i + kk] * B[kk * N + jj];
+                }
+                C[i * N + jj] = a;
+            }
+        }
+
+        A += GEMM_RM * K;
+        C += GEMM_RM * N;
+    }
+
+    // Tail rows: one at a time
+    for (; ii < M; ii++) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<1, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<1, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            float a = C[jj];
+            for (int64_t kk = 0; kk < K; kk++) {
+                a += A[kk] * B[kk * N + jj];
+            }
+            C[jj] = a;
+        }
+
+        A += K;
+        C += N;
+    }
+}
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
+#else // scalar path
+
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    for (int64_t i = 0; i < M; i++) {
+        for (int64_t j = 0; j < N; j++) {
+            float sum = C[i * N + j];
+            for (int64_t kk = 0; kk < K; kk++) {
+                sum += A[i * K + kk] * B[kk * N + j];
+            }
+            C[i * N + j] = sum;
+        }
+    }
+}
+
+#endif // GGML_SIMD

ggml/src/ggml-cpu/simd-mappings.h

@@ -1160,6 +1160,14 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     float32x4_t tmp = x[0] + vec_reve(x[0]);        \
     res = tmp[0] + tmp[1];                          \
 }
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
+{                                                \
+    float32x4_t v = vec_add(vec_add(s0, s1),     \
+                            vec_add(s2, s3));    \
+    v = vec_add(v, vec_sld(v, v, 8));            \
+    v = vec_add(v, vec_sld(v, v, 4));            \
+    res += (ggml_float)vec_extract(v, 0);        \
+}
 
 #define GGML_F32_VEC        GGML_F32x4
 #define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
@@ -1209,6 +1217,24 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_F16_VEC_MUL            GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
 
+// BF16 s390x
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR  8
+
+#define GGML_BF16x8         __vector unsigned short
+#define GGML_BF16x8_ZERO    vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC      GGML_BF16x8
+#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
+#define GGML_BF16_TO_F32_LO(v) ((float32x4_t) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((float32x4_t) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_FMA_LO(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
 #elif defined(__riscv_v_intrinsic)
 
 // compatible with vlen >= 128

ggml/src/ggml-cpu/vec.cpp

@@ -236,8 +236,7 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
     vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
     sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
 
-#endif
-#if defined(__POWER9_VECTOR__)
+#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
     const int np = (n & ~(GGML_BF16_STEP - 1));
     if (np > 0) {
         GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2872,6 +2872,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
     const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
     const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out";
     const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d";
+    const std::string delta_net_prefix = "dnet_add";
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -2902,7 +2903,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
             strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
             strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 &&
             strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 &&
-            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0) {
+            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0 &&
+            strncmp(node->name, delta_net_prefix.c_str(), delta_net_prefix.size()) != 0) {
             // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
             // by means of matching node names. See
             // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
@@ -4544,6 +4546,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
                 case GGML_UNARY_OP_TRUNC:
+                    // TODO: should become:
+                    //return ggml_is_contiguous_rows(op->src[0]);
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;

ggml/src/ggml-cuda/mmq.cuh

@@ -2715,14 +2715,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
 #pragma unroll
         for (int l = 0; l < QR2_XXS; ++l) {
-            const int * grid_pos = (const int *) (iq2xxs_grid + aux8[l]);
-            const int signs_packed = ksigns_iq2xs[(aux32 >> (7*l)) & 0x7F];
+            const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]];
+            const uint32_t signs = unpack_ksigns(aux32 >> (7 * l));
 
-            const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-            const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-            const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
@@ -2733,12 +2733,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
-        const int ls = aux32 >> 28;
+        const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
         const float d = bxi->d;
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
-        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = (ls*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #else
-        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4;
+        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 }
@@ -2776,11 +2776,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
     #pragma unroll
         for (int l = 0; l < QR2_XS; ++l) {
-            const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l] & 0x000001FF));
-            const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l] >> 9));
+            const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF];
+            const uint32_t signs = unpack_ksigns(q2[l] >> 9);
 
-            const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
@@ -2904,11 +2907,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #pragma unroll
         for (int l = 0; l < QR3_XXS; ++l) {
             const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
+            const uint32_t signs = unpack_ksigns(aux32 >> (7*l));
 
-            const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l)) & 0x7F));
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;

ggml/src/ggml-cuda/vecdotq.cuh

@@ -94,6 +94,15 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
 #endif
 }
 
+static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
+    // v is a 7 bit int, with the 8th sign being encodable as popcnt
+    // with xor we can "correct" the bit instead of having to mask
+    const uint32_t p = __popc(v) & 1;
+    const uint32_t s = v ^ p << 7;
+    // broadcast over uint to allow for 0x08040201 / 0x80402010 as selectors
+    return s * 0x01010101;
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
@@ -905,22 +914,22 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     int sumi = 0;
 #pragma unroll
     for (int k0 = 0; k0 < 8; k0 += 2) {
-        const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]);
-        const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F];
+        const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[k0/2]];
+        const uint32_t signs = unpack_ksigns(aux32 >> (7 * k0 / 2));
 
-        const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-        const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0);
         sumi = ggml_cuda_dp4a(grid0, u0, sumi);
 
-        const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-        const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1);
         sumi = ggml_cuda_dp4a(grid1, u1, sumi);
     }
 
-    const int ls = aux32 >> 28;
-    sumi = (ls*sumi + sumi/2)/4;
+    const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
+    sumi = sumi * ls / 8;           // (sumi * scale + sumi / 2) / 4
     const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
     return d * sumi;
 }
@@ -942,13 +951,15 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     int sumi1 = 0;
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
-        const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF));
-        const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l0/2] >> 9));
-
-        const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+        const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l0/2] & 0x1FF];
+        const uint32_t signs = unpack_ksigns(q2[l0/2] >> 9);
 
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         if (l0 < 4) {
@@ -1028,13 +1039,16 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
         const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]);
+        const uint32_t signs = unpack_ksigns(aux32 >> (7*l0/2));
 
-        const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F));
-
-        const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         sumi = ggml_cuda_dp4a(grid_l, u0, sumi);

ggml/src/ggml-metal/ggml-metal-common.cpp

@@ -273,6 +273,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             case GGML_OP_DIAG:
             case GGML_OP_MUL:
             case GGML_OP_ADD:
+            case GGML_OP_SUB:
             case GGML_OP_DIV:
             case GGML_OP_GLU:
             case GGML_OP_SCALE:

gguf-py/gguf/constants.py

@@ -3830,6 +3830,7 @@ class VisionProjectorType:
     MUSIC_FLAMINGO = "musicflamingo" # audio
     GLM4V = "glm4v"
     YOUTUVL = "youtuvl"
+    NEMOTRON_V2_VL = "nemotron_v2_vl"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -1346,6 +1346,7 @@ class TensorNameMap:
             "model.vision_tower.embeddings.cls_token", # Intern-S1
             "vision_model.class_embedding", # llama 4
             "model.vision.patch_embedding.cls_embedding", # cogvlm
+            "vision_model.radio_model.model.patch_generator.cls_token.token", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_PATCH: (
@@ -1360,6 +1361,7 @@ class TensorNameMap:
             "vision_tower.patch_embed.proj", # kimi-vl
             "model.vision.patch_embedding.proj", # cogvlm
             "siglip2.vision_model.embeddings.patch_embedding",
+            "vision_model.radio_model.model.patch_generator.embedder", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_NORM: (
@@ -1376,12 +1378,14 @@ class TensorNameMap:
             "visual.pos_embed", # qwen3vl
             "model.vision.patch_embedding.position_embedding", # cogvlm
             "visual.embeddings.position_embedding", # glm4v
+            "vision_model.radio_model.model.patch_generator.pos_embed", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
             "visual.blocks.{bid}.attn.qkv", # qwen3vl
             "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm
-            "vision_tower.encoder.blocks.{bid}.wqkv" # Kimi-K2.5
+            "vision_tower.encoder.blocks.{bid}.wqkv", # Kimi-K2.5
+            "vision_model.radio_model.model.blocks.{bid}.attn.qkv", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q: (
@@ -1446,6 +1450,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm1",
+            "vision_model.radio_model.model.blocks.{bid}.norm1", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1462,6 +1467,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.wo", # kimi-vl
             "model.vision.transformer.layers.{bid}.attention.dense", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl
+            "vision_model.radio_model.model.blocks.{bid}.attn.proj", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
@@ -1477,6 +1483,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm2",
+            "vision_model.radio_model.model.blocks.{bid}.norm2", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1493,6 +1500,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.mlp.fc1",
+            "vision_model.radio_model.model.blocks.{bid}.mlp.fc1", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1515,6 +1523,7 @@ class TensorNameMap:
             "vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm
             "siglip2.vision_model.encoder.layers.{bid}.mlp.fc2",
+            "vision_model.radio_model.model.blocks.{bid}.mlp.fc2", # Nemotron Nano v2 VL
         ),
 
         MODEL_TENSOR.V_LAYER_SCALE_1: (

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "f80864ca031932351abef49b74097c67f14719c6"
+HTTPLIB_VERSION = "d4180e923f846b44a3d30acd938438d6e64fc9f6"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

src/llama-context.cpp

@@ -878,6 +878,7 @@ const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) {
         }
     } catch (const std::exception & err) {
         // fallback to full vocab list
+        GGML_UNUSED(err);
     }
 
     return sampling.token_ids_full_vocab.data();
@@ -1809,7 +1810,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
 //
 
 uint32_t llama_context::output_reserve(int32_t n_outputs) {
-
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1893,11 +1893,6 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     embd = has_embd ? buffer_view<float>{(float *) (base + offset), embd.size} : buffer_view<float>{nullptr, 0};
     offset += embd.size * sizeof(float);
 
-    sampling.logits     = {nullptr, 0};
-    sampling.probs      = {nullptr, 0};
-    sampling.sampled    = {nullptr, 0};
-    sampling.candidates = {nullptr, 0};
-
     if (has_sampling) {
         sampling.logits = {(float *) (base + offset), (size_t)(n_vocab*n_outputs_max)};
         offset += sampling.logits.size * sizeof(float);
@@ -1923,6 +1918,15 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         std::fill(sampling.candidates_count.begin(), sampling.candidates_count.end(), 0);
 
         std::fill_n(sampling.sampled.data, sampling.sampled.size, LLAMA_TOKEN_NULL);
+    } else {
+        sampling.logits     = {nullptr, 0};
+        sampling.probs      = {nullptr, 0};
+        sampling.sampled    = {nullptr, 0};
+        sampling.candidates = {nullptr, 0};
+
+        sampling.logits_count.clear();
+        sampling.probs_count.clear();
+        sampling.candidates_count.clear();
     }
 
     // set all ids as invalid (negative)
@@ -1953,37 +1957,30 @@ void llama_context::output_reorder() {
             }
         }
 
-        if (sampling.logits.has_data()) {
+        if (!sampling.samplers.empty()) {
+            assert(sampling.logits.size > 0);
+            assert(sampling.probs.size > 0);
+            assert(sampling.candidates.size > 0);
+            assert(sampling.sampled.size > 0);
+            assert(sampling.logits_count.size() > 0);
+            assert(sampling.probs_count.size() > 0);
+            assert(sampling.candidates_count.size() > 0);
+
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.logits.data[i0*n_vocab + k], sampling.logits.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.probs.has_data()) {
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.probs.data[i0*n_vocab + k], sampling.probs.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.candidates.has_data()) {
             for (uint64_t k = 0; k < n_vocab; ++k) {
                 std::swap(sampling.candidates.data[i0*n_vocab + k], sampling.candidates.data[i1*n_vocab + k]);
             }
-        }
 
-        if (sampling.sampled.has_data()) {
-            std::swap(sampling.sampled.data[i0], sampling.sampled.data[i1]);
-        }
-
-        if (!sampling.logits_count.empty()) {
-            std::swap(sampling.logits_count[i0], sampling.logits_count[i1]);
-        }
-
-        if (!sampling.probs_count.empty()) {
-            std::swap(sampling.probs_count[i0], sampling.probs_count[i1]);
-        }
-
-        if (!sampling.candidates_count.empty()) {
+            std::swap(sampling.sampled.data[i0],     sampling.sampled.data[i1]);
+            std::swap(sampling.logits_count[i0],     sampling.logits_count[i1]);
+            std::swap(sampling.probs_count[i0],      sampling.probs_count[i1]);
             std::swap(sampling.candidates_count[i0], sampling.candidates_count[i1]);
         }
     }

src/llama-context.h

@@ -265,24 +265,26 @@ private:
     std::unique_ptr<llama_memory_i> memory;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
-    struct buffer_view<float>  logits = {nullptr, 0};
+    buffer_view<float> logits = {nullptr, 0};
 
     // embeddings output (2-dimensional array: [n_outputs][n_embd])
     // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
-    struct buffer_view<float>  embd = {nullptr, 0};
+    buffer_view<float> embd = {nullptr, 0};
 
     struct sampling_info {
+        // !samplers.empty() to check if any samplers are active
         std::map<llama_seq_id, llama_sampler *> samplers;
 
-        struct buffer_view<float>       logits     = {nullptr, 0};
-        struct buffer_view<llama_token> sampled    = {nullptr, 0};
-        struct buffer_view<float>       probs      = {nullptr, 0};
-        struct buffer_view<llama_token> candidates = {nullptr, 0};
+        buffer_view<float>       logits     = {nullptr, 0};
+        buffer_view<llama_token> sampled    = {nullptr, 0};
+        buffer_view<float>       probs      = {nullptr, 0};
+        buffer_view<llama_token> candidates = {nullptr, 0};
 
         std::vector<uint32_t> logits_count;
         std::vector<uint32_t> probs_count;
         std::vector<uint32_t> candidates_count;
 
+        // optimization
         std::vector<llama_token> token_ids_full_vocab;
     };
 

src/models/models.h

@@ -489,9 +489,6 @@ private:
     ggml_tensor * build_layer_attn_linear(
          llm_graph_input_rs * inp,
                 ggml_tensor * cur,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
@@ -506,9 +503,6 @@ private:
                 ggml_tensor * g,
                 ggml_tensor * beta,
                 ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
                         int   il);
 
     // returns pair of output and new state

src/models/qwen3next.cpp

@@ -16,17 +16,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
-                    GGML_TRI_TYPE_LOWER);
-
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
-    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
-
-    ggml_build_forward_expand(gf, causal_mask);
-    ggml_build_forward_expand(gf, identity);
-    ggml_build_forward_expand(gf, diag_mask);
-
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -36,7 +25,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
@@ -99,11 +88,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
         ggml_tensor * k,
         ggml_tensor * v,
         ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
-        ggml_tensor * diag_mask,
+        ggml_tensor * b,
+        ggml_tensor * s,
         int           il) {
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
@@ -113,134 +99,123 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
     const int64_t S_v = v->ne[0];
     const int64_t H_v = v->ne[1];
 
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
+    GGML_ASSERT(S_k == S_v);
+    GGML_ASSERT(H_v % H_k == 0);
 
     GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
     GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
 
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
 
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
+    const float scale = 1.0f / sqrtf(S_k);
 
     q = ggml_scale(ctx0, q, scale);
 
-    beta = ggml_sigmoid(ctx0, beta);
-
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
+    cb(b, "b_in", il);
     cb(g, "g_in", il);
 
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
+    g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [  1, n_tokens, H_v, n_seqs]
+    b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [  1, n_tokens, H_v, n_seqs]
 
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
+    const int CS = CHUNK_SIZE;
 
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
-
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
-
-    // Do padding
-    const int64_t chunk_size = CHUNK_SIZE;
-
-    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
-    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
+    const int pad = (CS - n_tokens % CS) % CS;
+    const int n_chunks = (n_tokens + pad) / CS;
 
     q = ggml_pad(ctx0, q, 0, pad, 0, 0);
     k = ggml_pad(ctx0, k, 0, pad, 0, 0);
     v = ggml_pad(ctx0, v, 0, pad, 0, 0);
-    g = ggml_pad(ctx0, g, pad, 0, 0, 0);
-    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
+    g = ggml_pad(ctx0, g, 0, pad, 0, 0);
+    b = ggml_pad(ctx0, b, 0, pad, 0, 0);
 
-    cb(q, "q_pad", il);
-    cb(k, "k_pad", il);
-    cb(v, "v_pad", il);
-    cb(beta, "beta_pad", il);
-    cb(g, "g_pad", il);
+    ggml_tensor * v_b = ggml_mul(ctx0, v, b);
+    ggml_tensor * k_b = ggml_mul(ctx0, k, b);
 
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
+    cb(v_b, "v_b", il);
+    cb(k_b, "k_b", il);
 
-    cb(v_beta, "v_beta", il);
-    cb(k_beta, "k_beta", il);
+    q   = ggml_reshape_4d(ctx0, q,   S_k, CS, n_chunks, H_k * n_seqs);
+    k   = ggml_reshape_4d(ctx0, k,   S_k, CS, n_chunks, H_k * n_seqs);
+    k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs);
+    v   = ggml_reshape_4d(ctx0, v,   S_v, CS, n_chunks, H_v * n_seqs);
+    v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
 
-    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
+    g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
 
-    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
+    cb(g_cs, "g_cs", il);
 
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
-    cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
+    ggml_tensor * g_cs_i = g_cs;
+    ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
 
-    ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
+    g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
 
-    ggml_tensor * gcs_j_broadcast =
-        ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-    cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
+    // [CS, CS, n_chunks, H_v * n_seqs]
+    ggml_tensor * decay_mask;
+    decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+    decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
     decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
+    cb(decay_mask, "decay_mask", il);
 
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * kb;
+    kb = ggml_mul_mat(ctx0, k,  k_b);
+    kb = ggml_mul    (ctx0, kb, decay_mask);
 
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-    cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * attn;
+    attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
 
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
+    ggml_tensor * identity;
+    identity = ggml_view_1d(ctx0, attn, CS, 0);
+    identity = ggml_fill   (ctx0, identity, 1.0f);
+    identity = ggml_diag   (ctx0, identity);
 
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-    cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
+    ggml_tensor * lhs = ggml_add(ctx0, attn, identity);
+    cb(lhs, "dnet_add_ch_lhs", il);
 
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
+    attn = ggml_neg(ctx0, attn);
 
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
+    ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
+    attn = ggml_add(ctx0, lin_solve, identity);
+    cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs]
 
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-    cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
+    // [S_v, CS, n_chunks, H_v * n_seqs]
+    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
 
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-    cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
 
-    ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
-    attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
-    attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
-    cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
+    k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
 
+    // [CS, S_k, n_chunks, H_k * n_seqs]
+    ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp);
+    cb(kbg, "k_beta_g_exp", il);
+
+    // [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
+    cb(k_cd, "k_cumdecay", il);
+
+    // [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp);
+    ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
+
+    // [CS, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+    kq = ggml_mul(ctx0, kq, decay_mask);
+    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
+    cb(kq, "kq", il);
 
     // vectorized calculation of key_gdiff
     // improved from the chunked version:
@@ -250,109 +225,98 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
     //   kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
     //   last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
 
-    // get last element in g_cumsum along chunk_size dimension (ne0)
+    // get last element in g_cumsum along CS dimension (ne0)
     // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
-                                        g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
-                                        (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
+    // [1, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
+            g_cs->nb[1],
+            g_cs->nb[2],
+            g_cs->nb[3],
+            ggml_row_size(g_cs->type, g_cs->ne[0] - 1));
+    cb(g_last, "g_last", il);
+
+    // TODO: remove this cont when CUDA supports non-cont unary ops
     g_last = ggml_cont(ctx0, g_last);
-    cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
 
+    // [1, 1, n_chunks, H_v * n_seqs]
     ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
+    cb(g_last_exp, "g_last_exp", il);
 
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
-    cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
+    // [CS, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
+    cb(g_diff, "g_diff", il);
 
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
-                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
+    ggml_tensor * g_diff_exp   = ggml_exp(ctx0, g_diff);
+    ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp);
 
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
-    cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
+    // [S_k, CS, n_chunks, H_v * n_seqs]
+    ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
+    cb(kg, "key_gdiff", il);
 
-    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
-    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
+    // [CS, S_k, n_chunks, H_v * n_seqs]
+    ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg));
+    cb(kg_t, "key_gdiff_t", il);
 
+    ggml_tensor * s_t = ggml_transpose(ctx0, s);
+    s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs);
+    cb(s_t, "dnet_add_ch_state", il);
 
-    // state to be updated per chunk
-    ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
-    cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
-
-    // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
-    ggml_tensor * core_attn_out = nullptr;
+    // [CS, S_v, n_chunks, H_v * n_seqs]
+    ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v));
 
     for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
-        // shape: (S_k, chunk_size, 1, H_k * n_seqs)
-        ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
+        ggml_tensor * ch_k_cd    = get_slice_2d(ctx0, k_cd,    chunk); // [S_k,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_v_t     = get_slice_2d(ctx0, v_t,     chunk); // [ CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * ch_kq      = get_slice_2d(ctx0, kq,      chunk); // [ CS,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k,  CS, 1, H_k * n_seqs]
+        ggml_tensor * ch_kg_t    = get_slice_2d(ctx0, kg_t,    chunk); // [ CS, S_k, 1, H_v * n_seqs]
 
-        // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
+        // [CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t);
+        cb(v_t_p, "v_prime", il);
 
-        // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-        ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
+        // [CS, S_v, 1, H_v * n_seqs]
+        ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p);
+        cb(v_t_new, "v_t_new", il);
 
-        // shape: (chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq);
+        cb(v_attn, "v_attn", il);
 
-        // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-        // replaced by precomputed attn_kq
-        ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
-        cb(attn_chunk, "attn_chunk", il);
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp);
+        cb(attn_inter, "attn_inter", il);
 
-        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
+        // [S_v, CS, 1, H_v * n_seqs]
+        ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn);
+        cb(o_ch, "dnet_add_ch_attn_out", il);
 
-        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
-        cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
-
-        // v_new = v_i - v_prime
-        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
-        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-        cb(v_new, "v_new_chunk", il);
-
-        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-        ggml_tensor * q_g_exp    = ggml_mul(ctx0, q_chunk, gexp_chunk);
-        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-        cb(attn_inter, "attn_inter_chunk", il);
-
-        // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
-        cb(v_attn, "v_attn_chunk", il);
-
-        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
-        cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-
-        core_attn_out = core_attn_out == nullptr
-            ? core_attn_out_chunk
-            : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
+        v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]);
 
         // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
-        //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
+        // TODO: head broadcast might not work here - probably will need a transpose
+        ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
 
         // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
-        new_state = ggml_add(ctx0,
-            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
-            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
+        ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
+        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
+        s_t = ggml_add(ctx0, s_t, kgv);
+        cb(s_t, "dnet_add_ch_state", il);
     }
 
+    s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs);
+
     // truncate padded tokens
-    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
+    ggml_tensor * o = ggml_view_4d(ctx0, v,
             S_v, n_tokens, H_v, n_seqs,
-            ggml_row_size(core_attn_out->type, S_v),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-    cb(output_tokens, "output_tokens", il);
+            ggml_row_size(v->type, S_v),
+            ggml_row_size(v->type, S_v * CS * n_chunks),
+            ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
 
-    // permute back to (S_v, H_v, n_tokens, n_seqs)
-    output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
-    output_tokens = ggml_cont(ctx0, output_tokens);
+    o = ggml_permute  (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
+    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
 
-    return {output_tokens, new_state};
+    return {o, s};
 }
 
 std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
@@ -360,8 +324,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_aut
         ggml_tensor * k,
         ggml_tensor * v,
         ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
+        ggml_tensor * b, // beta
+        ggml_tensor * s, // state
         int           il) {
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
@@ -371,75 +335,72 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_aut
     const int64_t S_v = v->ne[0];
     const int64_t H_v = v->ne[1];
 
-    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
+    GGML_ASSERT(n_tokens == 1);
+
+    GGML_ASSERT(S_k == S_v);
+    GGML_ASSERT(H_v % H_k == 0);
 
     GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
     GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
 
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
 
-    const float eps_norm = hparams.f_norm_rms_eps;
+    const float scale = 1.0f / sqrtf(S_k);
 
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
+    q = ggml_scale(ctx0, q, scale);
 
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q    = ggml_scale(ctx0, q, scale);
-    beta = ggml_sigmoid(ctx0, beta);
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
 
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
+    cb(b, "b_in", il);
     cb(g, "g_in", il);
 
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
+    g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
 
-    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
-    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
+    // [S_v, S_v, H_v, n_seqs]
+    g = ggml_exp(ctx0, g);
+    s = ggml_mul(ctx0, s, g);
 
-    // Apply exponential to g_t
-    g_t = ggml_exp(ctx0, g_t);
+    ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s));
 
-    // Apply the gated delta rule for the single timestep
-    // last_recurrent_state = last_recurrent_state * g_t
-    state = ggml_mul(ctx0, state, g_t);
+    // [1, S_v, H_v, n_seqs]
+    ggml_tensor * sk;
+    sk = ggml_mul     (ctx0, s_t, k);
+    sk = ggml_sum_rows(ctx0, sk);
 
-    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
-    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
-    // we need to sum over dim=-2, so we transpose, sum, then transpose again
-    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
+    // [S_v, 1, H_v, n_seqs]
+    ggml_tensor * d;
+    d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk));
+    d = ggml_mul(ctx0, d, b);
 
-    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
-    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    // delta = (v_t - kv_mem) * beta_t
-    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
-    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
+    // [1, S_v, H_v, n_seqs]
+    ggml_tensor * d_t;
+    d_t = ggml_transpose(ctx0, d);
 
-    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
-    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
-    state                   = ggml_add(ctx0, state, k_t_delta);
+    // [S_v, S_v, H_v, n_seqs]
+    ggml_tensor * kd;
+    k  = ggml_repeat(ctx0, k, s);
+    kd = ggml_mul   (ctx0, k, d_t);
 
-    // Compute the attention output
-    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
-    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
-    // again, since it's over dim = -2, transpose, sum, transpose back
-    ggml_tensor * core_attn_out =
-        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
+    s_t = ggml_add(ctx0, s_t, kd);
 
-    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
-    cb(core_attn_out, "output_tokens", il);
-    cb(state, "new_state", il);
+    cb(s_t, "dnet_add_ar_state", il);
 
-    return {core_attn_out, state};
+    ggml_tensor * s_q = ggml_mul     (ctx0, s_t, q);
+    ggml_tensor * o   = ggml_sum_rows(ctx0, s_q);
+
+    o = ggml_permute  (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
+    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
+
+    return {o, s};
 }
 
 ggml_tensor * llm_build_qwen3next::build_norm_gated(
@@ -472,39 +433,29 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
     // Split Q projection into query and gate
     // The split should be along dimension 0 (the feature dimension)
     ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
-                                             Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
+                                            Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
+    cb(Qcur, "Qcur_view", il);
+
     ggml_tensor * gate =
         ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
                      Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
-    cb(Qcur, "Qcur", il);
     cb(gate, "gate", il);
 
-    // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention
-    Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-    cb(Qcur, "Qcur_reshaped", il);
-
-    // Apply Q normalization
-    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
-    cb(Qcur, "Qcur_normed", il);
-
     ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
     cb(Kcur, "Kcur", il);
 
     ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
     cb(Vcur, "Vcur", il);
 
-    // Apply K normalization
     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Qcur, "Qcur_normed", il);
+
     Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
     cb(Kcur, "Kcur_normed", il);
 
-    // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads)
-    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
-    cb(gate, "gate_reshaped", il);
-
-    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-    // Apply RoPE
     Qcur = ggml_rope_ext(
             ctx0, Qcur, inp_pos, nullptr,
             n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -519,7 +470,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
     cb(Kcur, "Kcur", il);
     cb(Vcur, "Vcur", il);
 
-    // Attention computation
     const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
     cur = build_attn(inp,
@@ -527,10 +477,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
                 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
     cb(cur, "attn_pregate", il);
 
-    ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
-    cb(gate_sigmoid, "gate_sigmoid", il);
+    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
+    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
 
-    cur = ggml_mul(ctx0, cur, gate_sigmoid);
+    gate = ggml_sigmoid(ctx0, gate);
+    cb(gate, "gate_sigmoid", il);
+
+    gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
+
+    cur = ggml_mul(ctx0, cur, gate);
     cb(cur, "attn_gated", il);
 
     cur = build_lora_mm(model.layers[il].wo, cur);
@@ -560,7 +515,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
         cb(z, "z", il);
 
         return { qkv_mixed, z };
-
     } else {
         // legacy (slower) path
         ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);
@@ -624,9 +578,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
 ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
         llm_graph_input_rs * inp,
         ggml_tensor *        cur,
-        ggml_tensor *        causal_mask,
-        ggml_tensor *        identity,
-        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -671,7 +622,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
                                    split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));
     cb(a, "a", il);
 
-    ggml_tensor * beta  = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs);
+    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
+    b = ggml_cont(ctx0, b);
+
+    ggml_tensor * beta = ggml_sigmoid(ctx0, b);
+
+    beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
 
     // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
     ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
@@ -679,6 +635,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
+
     ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
     cb(gate, "gate", il);
 
@@ -686,8 +643,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
     ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
 
-    // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
-
     // Build the convolution states tensor
     ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     cb(conv_states, "conv_states", il);
@@ -696,11 +651,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;
     const int64_t conv_kernel_size = conv_kernel->ne[0];
     const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
-    conv_states                    = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
+
+    conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
     cb(conv_states, "conv_states_reshaped", il);
 
-    qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
-    cb(qkv_mixed, "qkv_mixed_permuted", il);
+    qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
+    cb(qkv_mixed, "qkv_mixed_transposed", il);
 
     ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
     cb(conv_input, "conv_input", il);
@@ -720,7 +676,10 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
     cb(conv_states_all, "conv_states_updated", il);
 
-    // Apply SSM convolution
+    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
+    cb(state, "state_predelta", il);
+
     ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
     cb(conv_output_proper, "conv_output_raw", il);
 
@@ -734,26 +693,36 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
 
     // Extract the convolved Q, K, V from conv_output
-    ggml_tensor * q_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
+    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            0);
+
+    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+
+    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_v_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
+
     cb(q_conv, "q_conv", il);
-    ggml_tensor * k_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(k_conv, "k_conv", il);
-    ggml_tensor * v_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(v_conv, "v_conv", il);
 
-    // Unsqueeze them
-    q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
-    state               = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
-    cb(state, "state_predelta", il);
+    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
+    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
+
+    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // if head keys and value keys are different, repeat to force tensors into matching shapes
     if (num_k_heads != num_v_heads) {
@@ -786,7 +755,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
     }
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
@@ -795,19 +764,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
 
     // Update the recurrent states
     ggml_build_forward_expand(gf,
-                              ggml_cpy(ctx0, new_state,
-                                       ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
-                                                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
-
-    // Reshape both attn_out_final and z to 2D tensors for normalization
-    // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+            ggml_cpy(ctx0, new_state,
+                ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
 
     // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // Apply gated normalization: self.norm(core_attn_out, z)
-    ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
+    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
 
     // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
     ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -818,7 +783,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(cur, "linear_attn_out", il);
 
     // Reshape back to original dimensions
-    cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+
     return cur;
 }
 
@@ -839,7 +805,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
         if (model.layers[il].ffn_up_shexp != nullptr) {
             ggml_tensor * ffn_shexp =
                 build_ffn(cur,
-                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
                     model.layers[il].ffn_gate_shexp, NULL, NULL,
                     model.layers[il].ffn_down_shexp, NULL, NULL,
                     NULL,
@@ -852,11 +818,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
             ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
             cb(shared_gate, "shared_expert_gate", il);
 
-            // Apply sigmoid to the gate
             shared_gate = ggml_sigmoid(ctx0, shared_gate);
             cb(shared_gate, "shared_expert_gate_sigmoid", il);
 
-            // Apply the gate to the shared expert output
             ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
             cb(ffn_shexp, "ffn_shexp_gated", il);
 

tests/test-backend-ops.cpp

@@ -8301,7 +8301,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                         //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
                                         for (int kv : { 113, 512, 1024, }) {
                                             if (nr2 != 1 && kv != 512) continue;
-                                            for (int nb : { 1, 3, 32, 35, }) {
+                                            for (int nb : { 1, 3, 32, 75, }) {
                                                 for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
                                                     if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
                                                     for (ggml_type type_KV : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {

tools/mtmd/CMakeLists.txt

@@ -20,6 +20,7 @@ add_library(mtmd
             models/internvl.cpp
             models/kimivl.cpp
             models/kimik25.cpp
+            models/nemotron-v2-vl.cpp
             models/llama4.cpp
             models/llava.cpp
             models/minicpmv.cpp

tools/mtmd/clip-impl.h

@@ -236,6 +236,7 @@ enum projector_type {
     PROJECTOR_TYPE_GLM4V,
     PROJECTOR_TYPE_YOUTUVL,
     PROJECTOR_TYPE_KIMIK25,
+    PROJECTOR_TYPE_NEMOTRON_V2_VL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -270,6 +271,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GLM4V,     "glm4v"},
     { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
     { PROJECTOR_TYPE_KIMIK25,   "kimik25"},
+    { PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip-model.h

@@ -15,6 +15,7 @@ enum ffn_op_type {
     FFN_GELU_ERF,
     FFN_SILU,
     FFN_GELU_QUICK,
+    FFN_RELU_SQR,
 };
 
 enum norm_type {

tools/mtmd/clip.cpp

@@ -559,6 +559,12 @@ ggml_tensor * clip_graph::build_ffn(
                 cur = ggml_gelu_quick(ctx0, cur);
                 cb(cur, "ffn_gelu_quick", il);
             } break;
+        case FFN_RELU_SQR:
+            {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_relu_sqr", il);
+            } break;
     }
 
     if (down) {
@@ -810,6 +816,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_internvl>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
+            {
+                builder = std::make_unique<clip_graph_nemotron_v2_vl>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_LLAMA4:
             {
                 builder = std::make_unique<clip_graph_llama4>(ctx, img);
@@ -1110,6 +1120,7 @@ struct clip_model_loader {
                         }
                     } break;
                 case PROJECTOR_TYPE_INTERNVL:
+                case PROJECTOR_TYPE_NEMOTRON_V2_VL:
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
@@ -1767,6 +1778,12 @@ struct clip_model_loader {
                     model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
                     model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
                 } break;
+            case PROJECTOR_TYPE_NEMOTRON_V2_VL:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
+                    model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
+                } break;
             case PROJECTOR_TYPE_GLMA:
                 {
                     model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -3088,6 +3105,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         case PROJECTOR_TYPE_GLM_EDGE:
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
             {
                 clip_image_u8 resized_image;
                 int sz = params.image_size;
@@ -3397,6 +3415,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
         case PROJECTOR_TYPE_LLAMA4:
             {
                 // both X and Y are downscaled by the scale factor
@@ -3805,6 +3824,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_GEMMA3NV:
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_GLMA:
         case PROJECTOR_TYPE_ULTRAVOX:
@@ -3968,6 +3988,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_NEMOTRON_V2_VL:
             return ctx->model.mm_3_w->ne[1];
         case PROJECTOR_TYPE_LLAMA4:
             return ctx->model.mm_model_proj->ne[1];

tools/mtmd/models/models.h

@@ -42,6 +42,11 @@ struct clip_graph_internvl : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_nemotron_v2_vl : clip_graph {
+    clip_graph_nemotron_v2_vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_llama4 : clip_graph {
     clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;

tools/mtmd/models/nemotron-v2-vl.cpp

@@ -0,0 +1,35 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_nemotron_v2_vl::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_registers = model.class_embedding->ne[1];
+    const int n_pos = n_patches + n_registers;
+
+    ggml_tensor * inp = build_inp();
+
+    // add position embeddings (pre-downsampled during GGUF conversion for fixed 512x512 input)
+    inp = ggml_add(ctx0, inp, model.position_embeddings);
+    cb(inp, "inp_pos", -1);
+
+    inp = ggml_concat(ctx0, model.class_embedding, inp, 1);
+
+    ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, nullptr);
+
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd),
+        n_registers * ggml_row_size(cur->type, n_embd));
+
+    cur = build_patch_merge_permute(cur, model.hparams.n_merge);
+
+    {
+        cur = build_norm(cur, model.mm_0_w, nullptr, NORM_TYPE_RMS, 1e-6, -1);
+        cur = build_ffn(cur, model.mm_1_w, nullptr, nullptr, nullptr, model.mm_3_w, nullptr, FFN_RELU_SQR, -1);
+    }
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/server/CMakeLists.txt

@@ -28,10 +28,6 @@ target_link_libraries(${TARGET} PUBLIC common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
 set(TARGET llama-server)
 
-if (NOT LLAMA_HTTPLIB)
-    message(FATAL_ERROR "LLAMA_HTTPLIB is OFF, cannot build llama-server. Hint: to skip building server, set -DLLAMA_BUILD_SERVER=OFF")
-endif()
-
 set(TARGET_SRCS
     server.cpp
     server-http.cpp

vendor/cpp-httplib/httplib.cpp

@@ -1264,78 +1264,32 @@ int poll_wrapper(struct pollfd *fds, nfds_t nfds, int timeout) {
 #endif
 }
 
-template <bool Read>
-ssize_t select_impl(socket_t sock, time_t sec, time_t usec) {
-#ifdef __APPLE__
-  if (sock >= FD_SETSIZE) { return -1; }
-
-  fd_set fds, *rfds, *wfds;
-  FD_ZERO(&fds);
-  FD_SET(sock, &fds);
-  rfds = (Read ? &fds : nullptr);
-  wfds = (Read ? nullptr : &fds);
-
-  timeval tv;
-  tv.tv_sec = static_cast<long>(sec);
-  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
-
-  return handle_EINTR([&]() {
-    return select(static_cast<int>(sock + 1), rfds, wfds, nullptr, &tv);
-  });
-#else
+ssize_t select_impl(socket_t sock, short events, time_t sec,
+                           time_t usec) {
   struct pollfd pfd;
   pfd.fd = sock;
-  pfd.events = (Read ? POLLIN : POLLOUT);
+  pfd.events = events;
+  pfd.revents = 0;
 
   auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
 
   return handle_EINTR([&]() { return poll_wrapper(&pfd, 1, timeout); });
-#endif
 }
 
 ssize_t select_read(socket_t sock, time_t sec, time_t usec) {
-  return select_impl<true>(sock, sec, usec);
+  return select_impl(sock, POLLIN, sec, usec);
 }
 
 ssize_t select_write(socket_t sock, time_t sec, time_t usec) {
-  return select_impl<false>(sock, sec, usec);
+  return select_impl(sock, POLLOUT, sec, usec);
 }
 
 Error wait_until_socket_is_ready(socket_t sock, time_t sec,
                                         time_t usec) {
-#ifdef __APPLE__
-  if (sock >= FD_SETSIZE) { return Error::Connection; }
-
-  fd_set fdsr, fdsw;
-  FD_ZERO(&fdsr);
-  FD_ZERO(&fdsw);
-  FD_SET(sock, &fdsr);
-  FD_SET(sock, &fdsw);
-
-  timeval tv;
-  tv.tv_sec = static_cast<long>(sec);
-  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
-
-  auto ret = handle_EINTR([&]() {
-    return select(static_cast<int>(sock + 1), &fdsr, &fdsw, nullptr, &tv);
-  });
-
-  if (ret == 0) { return Error::ConnectionTimeout; }
-
-  if (ret > 0 && (FD_ISSET(sock, &fdsr) || FD_ISSET(sock, &fdsw))) {
-    auto error = 0;
-    socklen_t len = sizeof(error);
-    auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR,
-                          reinterpret_cast<char *>(&error), &len);
-    auto successful = res >= 0 && !error;
-    return successful ? Error::Success : Error::Connection;
-  }
-
-  return Error::Connection;
-#else
   struct pollfd pfd_read;
   pfd_read.fd = sock;
   pfd_read.events = POLLIN | POLLOUT;
+  pfd_read.revents = 0;
 
   auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
 
@@ -1354,7 +1308,6 @@ Error wait_until_socket_is_ready(socket_t sock, time_t sec,
   }
 
   return Error::Connection;
-#endif
 }
 
 bool is_socket_alive(socket_t sock) {
@@ -7138,17 +7091,6 @@ Server::process_request(Stream &strm, const std::string &remote_addr,
   res.version = "HTTP/1.1";
   res.headers = default_headers_;
 
-#ifdef __APPLE__
-  // Socket file descriptor exceeded FD_SETSIZE...
-  if (strm.socket() >= FD_SETSIZE) {
-    Headers dummy;
-    detail::read_headers(strm, dummy);
-    res.status = StatusCode::InternalServerError_500;
-    output_error_log(Error::ExceedMaxSocketDescriptorCount, &req);
-    return write_response(strm, close_connection, req, res);
-  }
-#endif
-
   // Request line and headers
   if (!parse_request_line(line_reader.ptr(), req)) {
     res.status = StatusCode::BadRequest_400;
@@ -12063,7 +12005,7 @@ bool get_cert_sans(cert_t cert, std::vector<SanEntry> &sans) {
   if (!names) return true; // No SANs is valid
 
   auto count = sk_GENERAL_NAME_num(names);
-  for (int i = 0; i < count; i++) {
+  for (decltype(count) i = 0; i < count; i++) {
     auto gen = sk_GENERAL_NAME_value(names, i);
     if (!gen) continue;
 

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.31.0"
-#define CPPHTTPLIB_VERSION_NUM "0x001F00"
+#define CPPHTTPLIB_VERSION "0.32.0"
+#define CPPHTTPLIB_VERSION_NUM "0x002000"
 
 /*
  * Platform compatibility check