Sync master with upstream release b4966

ci/README.md

@@ -60,7 +60,7 @@ docker run --privileged -it \
 Inside the container, execute the following commands:
 
 ```bash
-apt update -y && apt install -y cmake git python3.10-venv wget
+apt update -y && apt install -y bc cmake git python3.10-venv time unzip wget
 git config --global --add safe.directory /ws
 GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
 ```

common/CMakeLists.txt

@@ -114,8 +114,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.6.12:
-        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
+        # v0.7.10:
+        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/llguidance.cpp

@@ -11,25 +11,24 @@ struct llama_sampler_llg {
     std::string         grammar_kind;
     std::string         grammar_data;
     LlgTokenizer *      tokenizer;
-    LlgConstraint *     grammar;
-    LlgMaskResult       llg_res;
-    bool                has_llg_res;
+    LlgMatcher *        grammar;
 };
 
-static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                             const char * grammar_data) {
+static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                          const char * grammar_data) {
     LlgConstraintInit cinit;
     llg_constraint_init_set_defaults(&cinit, tokenizer);
     const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
     if (log_level && *log_level) {
         cinit.log_stderr_level = atoi(log_level);
     }
-    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
-    if (llg_get_error(c)) {
-        LOG_ERR("llg error: %s\n", llg_get_error(c));
-        llg_free_constraint(c);
+    auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data);
+    if (llg_matcher_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_matcher_get_error(c));
+        llg_free_matcher(c);
         return nullptr;
     }
+
     return c;
 }
 
@@ -40,54 +39,39 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
 static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        LlgCommitResult res;
-        llg_commit_token(ctx->grammar, token, &res);
-        ctx->has_llg_res = false;
+        llg_matcher_consume_token(ctx->grammar, token);
     }
 }
 
 static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        if (!ctx->has_llg_res) {
-            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
-                ctx->has_llg_res = true;
+        const uint32_t * mask = llg_matcher_get_mask(ctx->grammar);
+        if (mask == nullptr) {
+            if (llg_matcher_compute_mask(ctx->grammar) == 0) {
+                mask = llg_matcher_get_mask(ctx->grammar);
             } else {
-                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
-                llg_free_constraint(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
+                llg_free_matcher(ctx->grammar);
                 ctx->grammar = nullptr;
+                return;
             }
         }
-        if (ctx->has_llg_res) {
-            if (ctx->llg_res.is_stop) {
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
-            } else {
-                const uint32_t * mask = ctx->llg_res.sample_mask;
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    auto token = cur_p->data[i].id;
-                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            auto token = cur_p->data[i].id;
+            if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                cur_p->data[i].logit = -INFINITY;
             }
         }
     }
 }
 
 static void llama_sampler_llg_reset(llama_sampler * smpl) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
-    if (!ctx->grammar) {
-        return;
+    if (ctx->grammar) {
+        llg_matcher_reset(ctx->grammar);
     }
-
-    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
-    llg_free_constraint(ctx->grammar);
-    ctx->grammar     = grammar_new;
-    ctx->has_llg_res = false;
 }
 
 static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
@@ -102,7 +86,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
         if (ctx->grammar) {
             result_ctx->grammar_kind = ctx->grammar_kind;
             result_ctx->grammar_data = ctx->grammar_data;
-            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
+            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
             result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
         }
     }
@@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_llg *) smpl->ctx;
 
     if (ctx->grammar) {
-        llg_free_constraint(ctx->grammar);
+        llg_free_matcher(ctx->grammar);
         llg_free_tokenizer(ctx->tokenizer);
     }
 
@@ -239,25 +223,24 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ grammar_data,
             /* .tokenizer    = */ tokenizer,
             /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
+        if (ctx->grammar) {
+            GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 ==
+                        llg_matcher_get_mask_byte_size(ctx->grammar));
+        }
     } else {
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_kind = */ {},
             /* .grammar_data = */ {},
             /* .tokenizer    = */ nullptr,
             /* .grammar      = */ nullptr,
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
     }
 
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx
-    );
+        /* .ctx   = */ ctx);
 }
 
 #else

convert_hf_to_gguf.py

@@ -1752,7 +1752,7 @@ class Mistral3Model(LlamaModel):
 
     # we need to merge the text_config into the root level of hparams
     def __init__(self, *args, **kwargs):
-        hparams = Model.load_hparams(kwargs["dir_model"])
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
         if "text_config" in hparams:
             hparams = {**hparams, **hparams["text_config"]}
             kwargs["hparams"] = hparams
@@ -3385,7 +3385,7 @@ class Gemma3Model(Model):
 
     # we need to merge the text_config into the root level of hparams
     def __init__(self, *args, **kwargs):
-        hparams = Model.load_hparams(kwargs["dir_model"])
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
         if "text_config" in hparams:
             hparams = {**hparams, **hparams["text_config"]}
             kwargs["hparams"] = hparams
@@ -3803,8 +3803,6 @@ def set_gguf_parameters(self):
     _tok_embd = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
         tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
 
@@ -3814,6 +3812,10 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             logger.debug("A_log --> A ==> " + new_name)
             data_torch = -torch.exp(data_torch)
 
+        # [4 1 8192 1] -> [4 8192 1 1]
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
         # assuming token_embd.weight is seen before output.weight
         if self._tok_embd is not None and new_name == output_name:
             if torch.equal(self._tok_embd, data_torch):
@@ -5358,7 +5360,7 @@ def main() -> None:
             logger.error(f"Model {model_architecture} is not supported")
             sys.exit(1)
 
-        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+        model_instance = model_class(dir_model, output_type, fname_out,
                                      is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
                                      eager=args.no_lazy,
                                      metadata_override=args.metadata, model_name=args.model_name,

docs/build.md

@@ -191,7 +191,7 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
@@ -218,6 +218,7 @@ By default, all supported compute capabilities are enabled. To customize this be
 
 ```bash
 cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
+cmake --build build --config Release
 ```
 
 This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.

examples/llava/clip.cpp

@@ -2989,7 +2989,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     assert(itype < GGML_TYPE_COUNT);
     ggml_type type = static_cast<ggml_type>(itype);
 
-    auto * ctx_clip = clip_model_load(fname_inp, 2);
+    auto * ctx_clip = clip_init(fname_inp, clip_context_params{
+        /* use_gpu */   false,
+        /* verbosity */ 2,
+    });
 
     const auto & ctx_src = ctx_clip->ctx_gguf;
     const auto & ctx_data = ctx_clip->ctx_data;