Merge pull request #193 from menloresearch/update-dev-from-master-2025-08-04-00-14

jan-service-account · web-flow · commit adb4fe282748 · 2025-08-04T07:27:23.000+07:00
Sync master with upstream release b6081
diff --git a/common/arg.cpp b/common/arg.cpp
@@ -2647,6 +2647,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_out_freq = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--output-format"}, "{gguf,dat}",
+        string_format("output format for imatrix file (default: %s)", params.imat_dat ? "dat" : "gguf"),
+        [](common_params & params, const std::string & value) {
+            /**/ if (value == "gguf") { params.imat_dat = false; }
+            else if (value == "dat")  { params.imat_dat = true;  }
+            else { throw std::invalid_argument("invalid output format"); }
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--save-frequency"}, "N",
         string_format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq),
diff --git a/common/common.h b/common/common.h
@@ -439,6 +439,7 @@ struct common_params {
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
+    bool    imat_dat    = false; // whether the legacy imatrix.dat format should be output
 
     bool process_output  = false; // collect data for the output tensor
     bool compute_ppl     = true;  // whether to compute perplexity
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
@@ -852,6 +852,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
             # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
             res = "exaone4"
+        if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
+            # ref: https://huggingface.co/JetBrains/Mellum-4b-base
+            res = "mellum"
 
         if res is None:
             logger.warning("\n")
@@ -6059,6 +6062,7 @@ def prepare_tensors(self):
 
 @ModelBase.register("DeepseekV2ForCausalLM")
 @ModelBase.register("DeepseekV3ForCausalLM")
+@ModelBase.register("KimiVLForConditionalGeneration")
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -6161,6 +6165,13 @@ def set_gguf_parameters(self):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip vision tensors and remove "language_model." for Kimi-VL
+        if "vision_tower" in name or "multi_modal_projector" in name:
+            return []
+
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+
         # rename e_score_correction_bias tensors
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
@@ -138,6 +138,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
+    {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -3096,6 +3096,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         uint32_t conv2d_SHMEM_PAD = 4;
         bool conv2d_UNROLL = true;
 
+        if (device->coopmat2) {
+            conv2d_SHMEM_PAD = 8; // 8 float16_t
+        }
+
         if (device->vendor_id == VK_VENDOR_ID_INTEL) {
             conv2d_SHMEM_PAD = 0;
             conv2d_UNROLL = false;
@@ -3154,7 +3158,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
         std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
         std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
 
-        if (conv2d_UNROLL) {
+        if (device->coopmat2) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else if (conv2d_UNROLL) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
                 sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
@@ -1,6 +1,11 @@
 #version 450
 
 #extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
 
 #ifdef USE_COLLECTIVES
 #    extension GL_KHR_shader_subgroup_shuffle : enable
@@ -91,6 +96,12 @@ uint32_t n_elems_out = K * NPQ;
 // Number of blocktiles per input
 uint32_t NB_CRS = splitWork(CRS, BS_CRS);
 
+#ifdef COOPMAT2
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
 const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
 const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
 
@@ -100,8 +111,8 @@ const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
 const uint32_t Ash_len = BS_K * Ash_stride;
 const uint32_t Bsh_len = BS_CRS * Bsh_stride;
 
-shared float Ash[Ash_len];  // K x CRS
-shared float Bsh[Bsh_len];  // CRS x NPQ
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
 
 // Threadtile sizes
 const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
@@ -110,10 +121,6 @@ const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
 const uint32_t NT_K   = BS_K / TS_K;
 const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
 
-float regA[TS_K];
-float regB[TS_NPQ];
-float regC[TS_K][TS_NPQ];
-
 /*
 Compute
 KxCRS @ CRSxNPQ = K x NPQ
@@ -145,12 +152,36 @@ uint fastdiv(uint n, uint mp, uint L) {
     return (msbs + n) >> L;
 }
 
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+    uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+    uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+    if (K_idx < K && NPQ_idx < NPQ) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
 void main() {
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#else
+    float regC[TS_K][TS_NPQ];
     for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
         for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
             regC[T_ly][T_lx] = 0.0;
         }
     }
+#endif
     /* Advance block in CRS dim */
     for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
         uint32_t CRS_idx_a;
@@ -199,7 +230,7 @@ void main() {
             if (K_idx >= K || CRS_idx_a >= CRS) {
                 val = 0.0;
             }
-            Ash[B_ly * Ash_stride + B_lx] = val;
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
         UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
@@ -244,11 +275,21 @@ void main() {
             if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
                 val = 0.0;
             }
-            Bsh[B_ly * Bsh_stride + B_lx] = val;
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
         }
         barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#else
         if (T_y * TS_K < K) {
             UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
                 for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
                     regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
                 }
@@ -262,9 +303,13 @@ void main() {
                 }
             }
         }
+#endif
         barrier();
     }
     /* Save C* */
+#ifdef COOPMAT2
+    coopMatPerElementNV(matC, matC, perElemOpStore);
+#else
     if (T_y * TS_K < K) {
         for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
             for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
@@ -280,4 +325,5 @@ void main() {
             }
         }
     }
+#endif
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -661,6 +661,9 @@ void process_shaders() {
     string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
     string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 
diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
@@ -312,7 +312,11 @@ def _try_load_from_config_json(self, path: Path) -> bool:
         with open(config_file, encoding = 'utf-8') as f:
             config = json.load(f)
         for typ in self.special_token_types:
-            self._set_special_token(typ, config.get(f'{typ}_token_id'))
+            token_id = config.get(f'{typ}_token_id')
+            # If not found at root, check in text_config (for multimodal models like Kimi-VL)
+            if token_id is None and 'text_config' in config:
+                token_id = config['text_config'].get(f'{typ}_token_id')
+            self._set_special_token(typ, token_id)
         return True
 
 
diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp
@@ -25,6 +25,7 @@ llama_memory_hybrid::llama_memory_hybrid(
                          /* common */
              uint32_t    n_seq_max,
                  bool    offload,
+                 bool    unified,
                          /* layer filters */
       layer_filter_cb && filter_attn,
       layer_filter_cb && filter_recr) :
@@ -38,7 +39,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
-        1,
+        unified,
         kv_size,
         n_seq_max,
         n_pad,
diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h
@@ -39,6 +39,7 @@ class llama_memory_hybrid : public llama_memory_i {
                              /* common */
                  uint32_t    n_seq_max,
                      bool    offload,
+                     bool    unified,
                              /* layer filters */
           layer_filter_cb && filter_attn = nullptr,
           layer_filter_cb && filter_recr = nullptr);
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
@@ -17598,6 +17598,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
                         /* n_seq_max         */ cparams.n_seq_max,
                         /* offload           */ cparams.offload_kqv,
+                        /* unified           */ cparams.kv_unified,
                         /* filter_attn       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr,
                         /* filter_recr       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr);
                 } else {
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
@@ -1856,7 +1856,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
-                    tokenizer_pre == "a.x-4.0") {
+                    tokenizer_pre == "a.x-4.0" ||
+                    tokenizer_pre == "mellum") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
diff --git a/tools/imatrix/README.md b/tools/imatrix/README.md
@@ -7,7 +7,7 @@ More information is available in <https://github.com/ggml-org/llama.cpp/pull/486
 
 ```
 ./llama-imatrix \
-    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \
+    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \
     [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \
     [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \
     [--show-statistics] [...]
@@ -20,6 +20,7 @@ The parameters in square brackets are optional and have the following meaning:
 * `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
 * `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
 * `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+* `--output-format` specifies the output format of the generated imatrix file. Either "gguf", or "dat" (the legacy format). Defaults to "gguf".
 * `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
 * `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
 * `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets.
@@ -45,14 +46,19 @@ Recent versions of `llama-imatrix` store data in GGUF format by default. For the
 
 ```bash
 # generate and save the imatrix using legacy format
-./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -o imatrix-legcy-format.dat -ngl 99
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --output-format dat -o imatrix-legcy-format.dat -ngl 99
 ```
 
 ```bash
-# covert legacy (binary) imatrix format to new (GGUF) format
+# convert legacy (binary) imatrix format to new (GGUF) format
 ./llama-imatrix --in-file imatrix-legacy-format.dat -o imatrix-new-format.gguf
 ```
 
+```bash
+# convert new (GGUF) imatrix format to legacy (binary) format
+./llama-imatrix --in-file imatrix-new-format.gguf --output-format dat -o imatrix-legacy-format.dat
+```
+
 ```bash
 # combine existing imatrices
 ./llama-imatrix --in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf -o imatrix-combined.gguf
diff --git a/tools/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp

Original file line number	Diff line number	Diff line change
`@@ -138,6 +138,7 @@ class TOKENIZER_TYPE(IntEnum):`
`138`	`138`	`{"name": "midm-2.0", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },`
`139`	`139`	`{"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},`
`140`	`140`	`{"name": "exaone4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },`
	`141`	`+ {"name": "mellum", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },`
`141`	`142`	`]`
`142`	`143`
`143`	`144`	`# some models are known to be broken upstream, so we will skip them as exceptions`