CUDA: tighter VRAM scratch size for 65b/70b (#2551)

* Update Vim plugin

* Remove getbufoneline usage, Add input bind example.

getbufoneline() appears to be a recently added function and has been
replaced with getbufline for compatibility.

An additional example that explains how to add a keybind that works in
insert mode was added.

examples/common.cpp

@@ -194,6 +194,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = 1.0f/std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -564,8 +570,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
     fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
     fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
     fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");

examples/llama.vim

@@ -0,0 +1,132 @@
+" Requires an already running llama.cpp server
+" To install either copy or symlink to ~/.vim/autoload/llama.vim
+" Then start with either :call llama#doLlamaGen(),
+" or add a keybind to your vimrc such as
+" nnoremap Z :call llama#doLlamaGen()<CR>
+" Similarly, you could add an insert mode keybind with
+" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
+"
+" g:llama_api_url and g:llama_overrides can be configured in your .vimrc
+" let g:llama_api_url = "192.168.1.10:8080"
+" llama_overrides can also be set through buffer/window scopes. For instance
+" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
+" Could be added to your .vimrc to automatically set a lower temperature when
+" editing a python script
+" Additionally, an override dict can be stored at the top of a file
+" !*{"stop": ["User:"]}
+" Could be added to the start of your chatlog.txt to set the stopping token
+" These parameter dicts are merged together from lowest to highest priority:
+" server default -> g:llama_overrides -> w:llama_overrides ->
+" b:llama_overrides -> in file (!*) overrides
+"
+" Sublists (like logit_bias and stop) are overridden, not merged
+" Example override:
+" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
+if !exists("g:llama_api_url")
+    let g:llama_api_url= "127.0.0.1:8080"
+endif
+if !exists("g:llama_overrides")
+   let g:llama_overrides = {}
+endif
+const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
+const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
+let s:linedict = {}
+
+func s:callbackHandler(bufn, channel, msg)
+   if len(a:msg) < 3
+      return
+   elseif a:msg[0] == "d"
+      let l:msg = a:msg[6:-1]
+   else
+      let l:msg = a:msg
+   endif
+   let l:decoded_msg = json_decode(l:msg)
+   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
+   if len(l:newtext) > 0
+      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
+   else
+      echo "nothing genned"
+   endif
+   if len(newtext) > 1
+      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
+      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
+   endif
+   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
+       echo "Finished generation"
+   endif
+endfunction
+
+func llama#doLlamaGen()
+   if exists("b:job")
+      if job_status(b:job) == "run"
+         call job_stop(b:job)
+         return
+      endif
+   endif
+
+   let l:cbuffer = bufnr("%")
+   let s:linedict[l:cbuffer] = line('$')
+   let l:buflines = getbufline(l:cbuffer, 1, 1000)
+   let l:querydata = copy(s:querydata)
+   call extend(l:querydata, g:llama_overrides)
+   if exists("w:llama_overrides")
+      call extend(l:querydata, w:llama_overrides)
+   endif
+   if exists("b:llama_overrides")
+      call extend(l:querydata, b:llama_overrides)
+   endif
+   if l:buflines[0][0:1] == '!*'
+      let l:userdata = json_decode(l:buflines[0][2:-1])
+      call extend(l:querydata, l:userdata)
+      let l:buflines = l:buflines[1:-1]
+   endif
+   let l:querydata.prompt = join(l:buflines, "\n")
+   let l:curlcommand = copy(s:curlcommand)
+   let l:curlcommand[2] = json_encode(l:querydata)
+   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
+endfunction
+
+" Echos the tokkenization of the provided string , or cursor to end of word
+" Onus is placed on the user to include the preceding space
+func llama#tokenizeWord(...)
+    if (a:0 > 0)
+        let l:input = a:1
+    else
+        exe "normal \"*ye"
+        let l:input = @*
+    endif
+    let l:querydata = {"content": l:input}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
+endfunction
+
+func s:tokenizeWordCallback(plaintext, channel, msg)
+    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
+endfunction
+
+
+" Echos the token count of the entire buffer (or provided string)
+" Example usage :echo llama#tokenCount()
+func llama#tokenCount(...)
+    if (a:0 > 0)
+        let l:buflines = a:1
+    else
+        let l:buflines = getline(1,1000)
+        if l:buflines[0][0:1] == '!*'
+            let l:buflines = l:buflines[1:-1]
+        endif
+        let l:buflines = join(l:buflines, "\n")
+    endif
+    let l:querydata = {"content": l:buflines}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
+endfunction
+
+func s:tokenCountCallback(channel, msg)
+    let resp = json_decode(a:msg)
+    echo len(resp.tokens)
+endfunction

examples/llm.vim

@@ -1,3 +1,5 @@
+" Basic plugin example
+
 function! Llm()
 
   let url = "http://127.0.0.1:8080/completion"
@@ -16,8 +18,10 @@ function! Llm()
   " Extract the content field from the response
   let content = json_decode(response).content
 
+  let split_newlines = split(content, '\n', 1)
+
   " Insert the content at the cursor position
-  call setline(line('.'), getline('.') . content)
+  call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
 endfunction
 
 command! Llm call Llm()

examples/main/README.md

@@ -140,6 +140,12 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 -   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
+### Extended Context Size
+
+Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
 ### Keep Prompt
 
 The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.

ggml.c

@@ -10731,72 +10731,96 @@ static void ggml_compute_forward_mul_mat(
         return;
     }
 
-    // parallelize by src0 rows
-    const int64_t dr = (ne01 + nth - 1)/nth;
-
-    const int64_t ir10 = dr*ith;
-    const int64_t ir11 = MIN(ir10 + dr, ne01);
-
-    // src1 rows
-    const int64_t nr1 = ne11*ne12*ne13;
-
     const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
     const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
-    for (int64_t ir1 = 0; ir1 < nr1; ++ir1) {
-        const int64_t i13 = (ir1/(ne12*ne11));
-        const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-        const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+    const int64_t nr0 = ne01;           // src0 rows
+    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
 
-        const int64_t ir0 = (ir1/ne11)%(ne02*ne03);
-        const int64_t i03 = (ir0/(ne02));
-        // Hack for "Falcon multi-query-attention key stutter" / alternative to ggml_repeat2.
-        // See https://github.com/ggerganov/llama.cpp/issues/1602#issuecomment-1606087470:
-        // GG: this is likely the correct way to broadcast, though need some more thought
-        //     therefore leaving the comments to remind us for now
-        const int64_t i02 = (i12 / (ne12 / ne02));
-        // Original from PR/224 (and also essential/correct for non-broadcast matmuls in Falcon)
-        // const int64_t i02 = (ir0 - i03*ne02);
+    //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
-        const int64_t i1 = i11;
-        const int64_t i2 = i12;
-        const int64_t i3 = i13;
+    // distribute the thread work across the inner or outer loop based on which one is larger
 
-        const char * src0_row = (const char *) src0->data + (  0 + i02*nb02 + i03*nb03     );
+    const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+    const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
 
-        // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
-        //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
-        //       the original src1 data pointer, so we should index using the indices directly
-        // TODO: this is a bit of a hack, we should probably have a better way to handle this
-        const char * src1_col = (const char *) wdata +
-            (src1_cont || src1->type != vec_dot_type
-             ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
-             : (i11*nb11 + i12*nb12 + i13*nb13));
+    const int64_t ith0 = ith % nth0;
+    const int64_t ith1 = ith / nth0;
 
-        float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+    const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+    const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
 
-        for (int64_t ir = ir10; ir < ir11; ++ir) {
-            vec_dot(ne00, &dst_col[ir], src0_row + ir*nb01, src1_col);
-        }
+    const int64_t ir010 = dr0*ith0;
+    const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+    const int64_t ir110 = dr1*ith1;
+    const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+    //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+    // threads with no work simply yield (not sure if it helps)
+    if (ir010 >= ir011 || ir110 >= ir111) {
+        sched_yield();
+        return;
     }
 
-    //int64_t t1 = ggml_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    printf("\n");
-    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
+    assert(ne12 % ne02 == 0);
+    assert(ne13 % ne03 == 0);
 
-    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    // block-tiling attempt
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
+
+    // attempt to reduce false-sharing (does not seem to make a difference)
+    float tmp[16];
+
+    for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+        for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                const int64_t i13 = (ir1/(ne12*ne11));
+                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
+                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+
+                // broadcast src0 into src1
+                const int64_t i03 = i13/r3;
+                const int64_t i02 = i12/r2;
+
+                const int64_t i1 = i11;
+                const int64_t i2 = i12;
+                const int64_t i3 = i13;
+
+                const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03);
+
+                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                //       the original src1 data pointer, so we should index using the indices directly
+                // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                     ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                     : (i11*nb11 + i12*nb12 + i13*nb13));
+
+                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                //}
+
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                }
+                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+            }
+        }
+    }
 }
 
-
 // ggml_compute_forward_out_prod
 
-
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,

llama.cpp

@@ -149,7 +149,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 }
 
 // amount of VRAM needed per batch size to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -157,14 +157,14 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
         { MODEL_7B,   512ull * kB },
         { MODEL_13B,  640ull * kB },
         { MODEL_30B,  768ull * kB },
-        { MODEL_65B, 1536ull * kB },
-        { MODEL_70B, 1536ull * kB }, // TODO (likely can be reduced)
+        { MODEL_65B, 1280ull * kB },
+        { MODEL_70B, 1280ull * kB },
     };
     return k_sizes;
 }
 
 // amount of VRAM needed per batch size and context to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -172,8 +172,8 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
         { MODEL_7B,  128ull },
         { MODEL_13B, 160ull },
         { MODEL_30B, 208ull },
-        { MODEL_65B, 416ull },
-        { MODEL_70B, 416ull }, // TODO (likely can be reduced)
+        { MODEL_65B, 256ull },
+        { MODEL_70B, 256ull },
     };
     return k_sizes;
 }