cuda: add SET operation support (#16804)

* feat(cuda): add GGML_OP_SET support

Implement CUDA kernel for SET operation with f32 support.

All tests passing (14598/14598).

* cuda(set): add I32 support; keep F32

* refactor(cuda): use ggml_cuda_cpy to unify SET operator logic and remove code duplication

* Update ggml/src/ggml-cuda/ggml-cuda.cu

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update ggml/src/ggml-cuda/set.cu

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

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

@@ -50,6 +50,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
@@ -2416,6 +2417,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -3842,6 +3846,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;

ggml/src/ggml-cuda/set.cu

@@ -0,0 +1,39 @@
+#include "set.cuh"
+#include "cpy.cuh"
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32));
+    GGML_ASSERT(src1->type == src0->type);
+    GGML_ASSERT(dst ->type == src0->type);
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const size_t nb1    = ((int32_t *) dst->op_params)[0];
+    const size_t nb2    = ((int32_t *) dst->op_params)[1];
+    const size_t nb3    = ((int32_t *) dst->op_params)[2];
+    const size_t offset = ((int32_t *) dst->op_params)[3];
+    const bool   inplace= (bool)     ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        ggml_cuda_cpy(ctx, src0, dst);
+    }
+
+    ggml_tensor dst_view = *dst;
+    dst_view.data  = (void *)((char *)dst->data + offset);
+    dst_view.ne[0] = src1->ne[0];
+    dst_view.ne[1] = src1->ne[1];
+    dst_view.ne[2] = src1->ne[2];
+    dst_view.ne[3] = src1->ne[3];
+
+    dst_view.nb[0] = ggml_element_size(dst);
+    dst_view.nb[1] = nb1;
+    dst_view.nb[2] = nb2;
+    dst_view.nb[3] = nb3;
+
+    ggml_cuda_cpy(ctx, src1, &dst_view);
+}

ggml/src/ggml-cuda/set.cuh

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_BLOCK_SIZE 256
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -211,7 +211,7 @@ static inline void hex_format_op_names(char * str, const struct ggml_tensor * t)
 // ** backend sessions
 
 struct ggml_hexagon_session {
-    ggml_hexagon_session(int dev_id) noexcept(false);
+    ggml_hexagon_session(int dev_id, ggml_backend_dev_t dev) noexcept(false);
     ~ggml_hexagon_session() noexcept(true);
 
     void allocate(int dev_id) noexcept(false);
@@ -1631,10 +1631,13 @@ void ggml_hexagon_session::release() noexcept(true) {
     }
 }
 
-ggml_hexagon_session::ggml_hexagon_session(int dev_id) noexcept(false) {
+ggml_hexagon_session::ggml_hexagon_session(int dev_id, ggml_backend_dev_t dev) noexcept(false) {
     buffer_type.context        = nullptr;
     repack_buffer_type.context = nullptr;
 
+    buffer_type.device         = dev;
+    repack_buffer_type.device  = dev;
+
     try {
         allocate(dev_id);
 
@@ -3628,7 +3631,7 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
         devices[i].iface   = ggml_backend_hexagon_device_i;
         devices[i].reg     = reg;
         try {
-            devices[i].context = new ggml_hexagon_session(i);
+            devices[i].context = new ggml_hexagon_session(i, &devices[i]);
         } catch (std::exception const &exc) {
             GGML_LOG_ERROR("ggml-hex: failed to create device/session %zu\n", i);
             devices[i].context = nullptr;

src/llama-kv-cache.cpp

@@ -961,10 +961,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;
@@ -2014,8 +2018,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
-
-uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
-    // the FA kernels require padding to avoid extra runtime boundary checks
-    return cparams.flash_attn ? 256u : 32u;
-}

src/llama-kv-cache.h

@@ -19,8 +19,6 @@ struct llama_context;
 
 class llama_kv_cache : public llama_memory_i {
 public:
-    static uint32_t get_padding(const llama_cparams & cparams);
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());

src/llama-model.cpp

@@ -19641,7 +19641,7 @@ struct llm_build_apertus : public llm_graph_context {
     }
 };
 
-llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
+llama_memory_i * llama_model::create_memory(const llama_memory_params & params, const llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
@@ -19692,17 +19692,13 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         };
                     }
 
-                    const auto padding = llama_kv_cache::get_padding(cparams);
-
-                    cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
-
                     res = new llama_memory_hybrid(
                         /* model             */ *this,
                         /* attn_type_k       */ params.type_k,
                         /* attn_type_v       */ params.type_v,
                         /* attn_v_trans      */ !cparams.flash_attn,
                         /* attn_kv_size      */ cparams.n_ctx,
-                        /* attn_n_pad        */ padding,
+                        /* attn_n_pad        */ 1,
                         /* attn_n_swa        */ hparams.n_swa,
                         /* attn_swa_type     */ hparams.swa_type,
                         /* recurrent_type_k  */ GGML_TYPE_F32,
@@ -19714,23 +19710,12 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* filter_attn       */ std::move(filter_attn),
                         /* filter_recr       */ std::move(filter_recr));
                 } else {
-                    const auto padding = llama_kv_cache::get_padding(cparams);
-
                     uint32_t n_ctx_per_stream = cparams.n_ctx;
 
                     if (!cparams.kv_unified) {
                         n_ctx_per_stream = (cparams.n_ctx + cparams.n_seq_max - 1)/cparams.n_seq_max;
-                        n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
-
-                        cparams.n_ctx = n_ctx_per_stream*cparams.n_seq_max;
-                    } else {
-                        n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
-
-                        cparams.n_ctx = n_ctx_per_stream;
                     }
 
-                    LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
-
                     llama_memory_i::layer_reuse_cb reuse = nullptr;
 
                     if (arch == LLM_ARCH_GEMMA3N) {
@@ -19757,7 +19742,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 n_ctx_per_stream,
                                 cparams.n_seq_max,
                                 cparams.n_ubatch,
-                                padding,
+                                1,
                                 nullptr,
                                 reuse);
                     } else {
@@ -19772,7 +19757,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 cparams.kv_unified,
                                 n_ctx_per_stream,
                                 cparams.n_seq_max,
-                                padding,
+                                1,
                                 hparams.n_swa,
                                 hparams.swa_type,
                                 nullptr,

src/llama-model.h

@@ -500,9 +500,8 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;

tools/llama-bench/README.md

@@ -82,6 +82,9 @@ Using the `-d <n>` option, each test can be run at a specified context depth, pr
 
 For a description of the other options, see the [main example](../main/README.md).
 
+> [!NOTE]
+> The measurements with `llama-bench` do not include the times for tokenization and for sampling.
+
 ## Examples
 
 ### Text generation with different models
@@ -131,7 +134,7 @@ $ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
-| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 |
 
 ### Different numbers of layers offloaded to the GPU
 

tools/server/server.cpp

@@ -2866,10 +2866,12 @@ struct server_context {
 
         // if context shifting is disabled, make sure that we don't run out of context
         if (!params_base.ctx_shift && slot.n_past + 1 >= slot.n_ctx) {
+            slot.truncated      = true;
             slot.stop           = STOP_TYPE_LIMIT;
             slot.has_next_token = false;
 
-            SLT_DBG(slot, "stopped due to running out of context, n_past = %d, n_ctx = %d\n", slot.n_past, slot.n_ctx);
+            SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n",
+                    slot.n_decoded, slot.n_prompt_tokens(), slot.n_past, slot.n_ctx);
         }
 
         // check the limits
@@ -2929,16 +2931,6 @@ struct server_context {
             }
         }
 
-        // if context shift is disabled, we stop when it reaches the context limit
-        if (slot.n_past >= slot.n_ctx) {
-            slot.truncated      = true;
-            slot.stop           = STOP_TYPE_LIMIT;
-            slot.has_next_token = false;
-
-            SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n",
-                    slot.n_decoded, slot.n_prompt_tokens(), slot.n_past, slot.n_ctx);
-        }
-
         if (llama_vocab_is_eog(vocab, result.tok)) {
             slot.stop           = STOP_TYPE_EOS;
             slot.has_next_token = false;
@@ -2946,19 +2938,6 @@ struct server_context {
             SLT_DBG(slot, "%s", "stopped by EOS\n");
         }
 
-        const auto n_ctx_train = llama_model_n_ctx_train(model);
-
-        if (slot.task->params.n_predict < 1 && slot.n_prompt_tokens() + slot.n_decoded >= n_ctx_train) {
-            slot.truncated      = true;
-            slot.stop           = STOP_TYPE_LIMIT;
-            slot.has_next_token = false; // stop prediction
-
-            SLT_WRN(slot,
-                    "n_predict (%d) is set for infinite generation. "
-                    "Limiting generated tokens to n_ctx_train (%d) to avoid EOS-less generation infinite loop\n",
-                    slot.task->params.n_predict, n_ctx_train);
-        }
-
         SLT_DBG(slot, "n_decoded = %d, n_remaining = %d, next token: %5d '%s'\n", slot.n_decoded, slot.n_remaining, result.tok, token_str.c_str());
 
         return slot.has_next_token; // continue

tools/server/tests/unit/test_ctx_shift.py

@@ -45,7 +45,7 @@ def test_ctx_shift_enabled():
 
 @pytest.mark.parametrize("n_predict,n_token_output,truncated", [
     (64, 64, False),
-    (-1, 120, True),
+    (-1, 248, True), # 8 tokens prompt + 248 tokens generated = 256 tokens total
 ])
 def test_ctx_shift_disabled_short_prompt(n_predict: int, n_token_output: int, truncated: bool):
     global server