Merge pull request #129 from ggerganov/0cc4m/vulkan-embedding-fix

0cc4m/vulkan embedding fix

.github/labeler.yml

@@ -0,0 +1,73 @@
+# https://github.com/actions/labeler
+
+SYCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-sycl.h
+            - ggml-sycl.cpp
+            - README-sycl.md
+Nvidia GPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-cuda/**
+Vulkan:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml_vk_generate_shaders.py
+            - ggml-vulkan*
+documentation:
+    - changed-files:
+        - any-glob-to-any-file:
+            - docs/**
+            - media/**
+testing:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tests/**
+build:
+    - changed-files:
+        - any-glob-to-any-file:
+            - cmake/**
+            - CMakeLists.txt
+            - CMakePresets.json
+            - codecov.yml
+examples:
+    - changed-files:
+        - any-glob-to-any-file: examples/**
+devops:
+    - changed-files:
+        - any-glob-to-any-file:
+            - .devops/**
+            - .github/**
+            - ci/**
+python:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.py"
+            - requirements/**
+            - gguf-py/**
+            - .flake8
+script:
+    - changed-files:
+        - any-glob-to-any-file:
+            - scripts/**
+android:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/llama.android/**
+server:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/server/**
+ggml:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-*.c
+            - ggml-*.h
+            - ggml-cuda/**
+nix:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.nix"
+            - .github/workflows/nix-*.yml
+            - .devops/nix/nixpkgs-instances.nix

.github/workflows/labeler.yml

@@ -0,0 +1,12 @@
+name: "Pull Request Labeler"
+on:
+- pull_request_target
+
+jobs:
+  labeler:
+    permissions:
+      contents: read
+      pull-requests: write
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/labeler@v5

CMakeLists.txt

@@ -581,7 +581,7 @@ if (LLAMA_HIPBLAS)
     else()
         # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
         if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_ARGETS})
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
         endif()
         cmake_minimum_required(VERSION 3.21)
         enable_language(HIP)

convert-hf-to-gguf.py

@@ -573,6 +573,10 @@ def _set_vocab_sentencepiece(self):
 
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
         for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
@@ -588,21 +592,23 @@ def _set_vocab_sentencepiece(self):
             elif tokenizer.IsByte(token_id):
                 toktype = SentencePieceTokenTypes.BYTE
 
-            tokens.append(text)
-            scores.append(score)
-            toktypes.append(toktype)
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
 
         added_tokens_file = self.dir_model / 'added_tokens.json'
         if added_tokens_file.is_file():
             with open(added_tokens_file, "r", encoding="utf-8") as f:
                 added_tokens_json = json.load(f)
-
                 for key in added_tokens_json:
-                    key = key.encode("utf-8")
-                    if key not in tokens:
-                        tokens.append(key)
-                        scores.append(-1000.0)
-                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    token_id = added_tokens_json[key]
+                    if (token_id >= vocab_size):
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
         if vocab_size > len(tokens):
             pad_count = vocab_size - len(tokens)
@@ -612,8 +618,6 @@ def _set_vocab_sentencepiece(self):
                 scores.append(-1000.0)
                 toktypes.append(SentencePieceTokenTypes.UNUSED)
 
-        assert len(tokens) == vocab_size
-
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)

ggml-quants.c

@@ -14,6 +14,12 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT
 
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid warnings for hundreds of casts
 // we should just be careful :)
@@ -1109,7 +1115,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (amax < 1e-30f) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -1177,7 +1183,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) { L[i] = 0; }
         return 0.f;
     }
@@ -1646,7 +1652,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
             break;
         }
     }
-    return sumlx / suml2;
+    return sumlx/suml2;
 }
 
 static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
@@ -2653,7 +2659,7 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -2805,7 +2811,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -12599,7 +12605,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -12775,7 +12781,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 16);
                 continue;
@@ -13216,7 +13222,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ3_XXS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -13756,7 +13762,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_S) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -13944,7 +13950,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
             }
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_M) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -14208,7 +14214,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
                 amax = ax; max = xb[j];
             }
         }
-        if (!amax) {
+        if (amax < GROUP_MAX_EPS) {
             scales[ib] = 0;
             continue;
         }
@@ -14429,7 +14435,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ2_S) {
                 scales[ib] = 0;
                 continue;
             }