chat : fix muse-glimmer detection of tool calls after EOM (#26879)

* chat : fix muse-glimmer swallowing a trailing tool call into content

Muse Glimmer routinely answers the user and calls a tool in a single
generation. The template terminates a message with <|eom|> when more
messages follow in the same turn and <|eot|> only at the end of the turn,
so the answer is closed by <|eom|> and the call opens a fresh header:

    <prose><|eom|><|start|>assistant to=<tool><|message|><atem:function_calls>...

The final-message rule read content with until("<|eot|>"), which assumed the
user-facing message is always last. There is no <|eot|> before the call, so
content ran to the end of the turn, absorbed the markup, and no tool_calls
were emitted - the tool never ran. On a tau2-bench telecom run this hit 43
turns across 19 of 114 tasks.

Stop the answer at <|eom|> and parse what follows as tool calls.

Adds models/templates/muse-glimmer.jinja and four parser tests: a plain
answer, the <|eom|> junction, markup quoted in an answer staying content,
and tool markup inside the to=self channel staying reasoning.

* address comment

.github/workflows/release.yml

@@ -749,6 +749,9 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows-rocm:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
     runs-on: windows-2022
 
     strategy:

.github/workflows/server-sanitize.yml

@@ -110,7 +110,7 @@ jobs:
           source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
@@ -119,4 +119,4 @@ jobs:
           source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh

.github/workflows/server-self-hosted.yml

@@ -72,7 +72,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +99,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
   server-cuda:
     runs-on: [self-hosted, llama-server, Linux, NVIDIA]
@@ -132,7 +132,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -141,7 +141,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -150,7 +150,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -159,7 +159,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
   server-kleidiai:
     runs-on: ah-ubuntu_22_04-c8g_8x
@@ -219,4 +219,4 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          pytest -v -x -m "not slow"
+          ./tests.sh

.github/workflows/server.yml

@@ -104,21 +104,21 @@ jobs:
         id: server_integration_tests
         run: |
           cd tools/server/tests
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
         run: |
           cd tools/server/tests
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh
 
       - name: Tests (Backend sampling)
         id: server_integration_tests_backend_sampling
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          pytest -v -x -m "not slow"
+          ./tests.sh
 
       - name: Slow tests (Backend sampling)
         id: server_integration_tests_slow_backend_sampling
@@ -126,7 +126,7 @@ jobs:
         run: |
           cd tools/server/tests
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          SLOW_TESTS=1 pytest -v -x
+          SLOW_TESTS=1 ./tests.sh
 
   windows:
     runs-on: windows-2025
@@ -167,15 +167,17 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
+        shell: bash
         run: |
           cd tools/server/tests
-          $env:PYTHONIOENCODING = ":replace"
-          pytest -v -x -m "not slow"
+          export PYTHONIOENCODING=":replace"
+          ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ github.event.schedule || github.event.inputs.slow_tests == 'true' }}
+        shell: bash
         run: |
           cd tools/server/tests
-          $env:SLOW_TESTS = "1"
-          pytest -v -x
+          export SLOW_TESTS="1"
+          ./tests.sh

ci/run.sh

@@ -49,6 +49,14 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
+# gpu-rocm self-hosted runner can't upload logs to blob; keep each run's logs in
+# their own dir keyed by the GitHub run id so an Actions run URL maps to its logs.
+if [ -n "${GG_BUILD_ROCM}" ] && [ -n "${GITHUB_RUN_ID}" ]; then
+    OUT="$OUT/run-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
+    mkdir -p "$OUT"
+    echo "ci results dir: $OUT"
+fi
+
 rm -f $OUT/*.log
 rm -f $OUT/*.exit
 rm -f $OUT/*.md

common/arg.cpp

@@ -2605,14 +2605,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_DIO"));
     add_opt(common_arg(
         {"-lm", "--load-mode"}, "MODE",
-        "model loading mode (default: mmap)\n"
+        "model loading mode (default: auto)\n"
+        "- auto: mmap, unless a device does not support it\n"
         "- none: no special loading mode\n"
         "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
         "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
         "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
         "- dio: use DirectIO if available\n",
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            /**/ if (value == "auto")       { params.load_mode = LLAMA_LOAD_MODE_AUTO;       }
+            else if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
             else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
             else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
             else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }

common/chat.cpp

@@ -1166,6 +1166,16 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
         data.prompt += data.generation_prompt;
     }
 
+    std::vector<std::string> tool_call_starts = { "<tool_call>" };
+
+    // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
+    // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
+    // constraining on <function which may occur in valid content generation, e.g. #include <functional>
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const std::string name = tool.at("function").at("name");
+        tool_call_starts.push_back("<function=" + name + ">");
+    });
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto generation_prompt = p.literal(GEN_PREFIX);
 
@@ -1238,7 +1248,7 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
             auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
 
             return generation_prompt +
-                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
+                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
         }
 
         // Content only parser
@@ -1264,12 +1274,9 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
         });
 
         if (data.grammar_lazy) {
-            data.grammar_triggers = {
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
-                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
-                // the token with the function name e.g. "=read"
-                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
-            };
+            for (const auto & start : tool_call_starts) {
+                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
+            }
         }
     }
 
@@ -3148,7 +3155,8 @@ static common_chat_params common_chat_params_init_muse_glimmer(const common_chat
         auto analysis = p.ref("analysis");
 
         auto recipient  = p.optional(p.literal(" to=user"));
-        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") + p.content(p.until("<|eot|>")));
+        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
+                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));
 
         if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
             auto string_value = p.ac(
@@ -3204,7 +3212,8 @@ static common_chat_params common_chat_params_init_muse_glimmer(const common_chat
             if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
                 return p.zero_or_more(start + analysis) + start + tool_calls;
             }
-            return p.zero_or_more(start + analysis) + start + (tool_calls | final_msg);
+            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
+            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
         }
 
         return p.zero_or_more(start + analysis) + start + final_msg;

common/common.h

@@ -473,7 +473,7 @@ struct common_params {
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
+    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
 
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;

common/peg-parser.cpp

@@ -570,23 +570,34 @@ struct parser_executor {
     }
 
     static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
+        auto save = pos;
+
         ++pos; // consume '\'
         if (pos >= ctx.input.size()) {
             if (!ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
             }
+            pos = save; // suppress unmatched '\'
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
         }
 
         char c = ctx.input[pos];
+
         if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
             ++pos;
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
-        } else if (c == 'u') {
-            return handle_unicode_escape(ctx, start, pos);
-        } else {
-            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
         }
+
+        if (c == 'u') {
+            auto result = handle_unicode_escape(ctx, start, pos);
+            if (result.need_more_input()) {
+                pos = save; // suppress incomplete sequence
+                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
+            }
+            return result;
+        }
+
+        return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
     }
 
     static common_peg_parse_result handle_unicode_escape(common_peg_parse_context & ctx, size_t start, size_t & pos) {

common/speculative.cpp

@@ -171,12 +171,6 @@ struct common_speculative_impl {
     // (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
     virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
     virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}
-
-    // true if this implementation requires the target context to extract post-norm embeddings
-    virtual bool need_embd() const = 0;
-
-    // true if this implementation requires the target context to extract pre-norm embeddings
-    virtual bool need_embd_nextn() const { return false; }
 };
 
 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -193,6 +187,10 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         auto * ctx_tgt = this->params.ctx_tgt;
 
+        if (!ctx_dft) {
+            throw std::runtime_error("draft-simple requires a draft context");
+        }
+
         SPC_TRC("%s", "adding speculative implementation 'draft-simple'\n");
         SPC_TRC("- n_max=%d, n_min=%d, p_min=%f\n", this->params.n_max, this->params.n_min, this->params.p_min);
         SPC_TRC("- gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n",
@@ -385,10 +383,6 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 
@@ -907,10 +901,6 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         pending_g_last[seq_id].resize(n_embd_dec);
         std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 // DFlash: block-diffusion drafting with a draft-side KV cache injection
@@ -1247,10 +1237,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 struct common_speculative_impl_draft_mtp : public common_speculative_impl {
@@ -1689,14 +1675,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
         std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
     }
-
-    bool need_embd() const override {
-        return false;
-    }
-
-    bool need_embd_nextn() const override {
-        return true;
-    }
 };
 
 // state of self-speculation (simple implementation, not ngram-map)
@@ -1743,10 +1721,6 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
@@ -1801,10 +1775,6 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
 
         common_ngram_map_accept(config[seq_id], n_accepted);
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 struct common_speculative_impl_ngram_mod : public common_speculative_impl {
@@ -1980,10 +1950,6 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
             }
         }
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 struct common_speculative_impl_ngram_cache : public common_speculative_impl {
@@ -2123,10 +2089,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
     void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
-
-    bool need_embd() const override {
-        return false;
-    }
 };
 
 struct common_speculative {
@@ -2322,7 +2284,6 @@ common_speculative_init_result::common_speculative_init_result(
     const bool spec_mtp = std::find(params.speculative.types.begin(),
                                     params.speculative.types.end(),
                                     COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-    GGML_ASSERT(has_draft || spec_mtp);
 
     auto mparams = common_model_params_to_llama(params);
     auto cparams = common_context_params_to_llama(params);
@@ -2560,34 +2521,6 @@ bool common_speculative_process(common_speculative * spec, const llama_batch & b
     return result;
 }
 
-bool common_speculative_need_embd(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
-bool common_speculative_need_embd_nextn(common_speculative * spec) {
-    if (spec == nullptr) {
-        return false;
-    }
-
-    for (auto & impl : spec->impls) {
-        if (impl->need_embd_nextn()) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
 void common_speculative_draft(common_speculative * spec) {
     if (spec == nullptr) {
         return;
@@ -2672,7 +2605,10 @@ void common_speculative_draft(common_speculative * spec) {
 void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, uint16_t n_accepted) {
     common_speculative_impl * impl = spec->impl_last[seq_id];
 
-    GGML_ASSERT(impl);
+    if (impl == nullptr) {
+        GGML_ASSERT(n_accepted == 0);
+        return;
+    }
 
     {
         common_time_meas tm(impl->t_accept_us, !impl->gen_perf);

common/speculative.h

@@ -67,12 +67,6 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);
 
-// true if any implementation requires target post-norm embeddings to be extracted
-bool common_speculative_need_embd(common_speculative * spec);
-
-// true if any implementation requires target nextn embeddings to be extracted
-bool common_speculative_need_embd_nextn(common_speculative * spec);
-
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);
 

conversion/__init__.py

@@ -214,6 +214,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Qwen3MoeForCausalLM": "qwen",
     "Qwen3NextForCausalLM": "qwen",
     "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
     "Qwen3TTSForConditionalGeneration": "qwen3tts",
     "Qwen3VLForConditionalGeneration": "qwen3vl",
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
@@ -310,6 +311,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Qwen2_5_VLForConditionalGeneration": "qwenvl",
     "Qwen3ASRForConditionalGeneration": "qwen3vl",
     "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "PocketTTSModel": "pockettts",
     "Qwen3TTSForConditionalGeneration": "qwen3tts",
     "Qwen3VLForConditionalGeneration": "qwen3vl",
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",

conversion/base.py

@@ -58,6 +58,11 @@ logger = logging.getLogger("hf-to-gguf")
 AnyModel = TypeVar("AnyModel", bound="type[ModelBase]")
 
 
+# for checkpoints that ship no config.json, we will try to provide a synthetic one
+HparamsMatcher = Callable[[Path], bool]
+HparamsLoader = Callable[[Path], dict[str, Any]]
+
+
 class SentencePieceTokenTypes(IntEnum):
     NORMAL = 1
     UNKNOWN = 2
@@ -77,6 +82,7 @@ class ModelBase:
         ModelType.TEXT: {},
         ModelType.MMPROJ: {},
     }
+    _hparams_loaders: list[tuple[HparamsMatcher, HparamsLoader]] = []
 
     dir_model: Path
     ftype: gguf.LlamaFileType
@@ -823,7 +829,7 @@ class ModelBase:
             elif any(str(v.get("quant_algo")).endswith("NVFP4") for v in quant_layers.values() if isinstance(v, dict)):
                 quant_algo = "NVFP4"
 
-        self._is_nvfp4 = quant_algo == "NVFP4"
+        self._is_nvfp4 = quant_algo in ("NVFP4", "W4A16_NVFP4")
         self._is_mxfp4 = quant_method == "mxfp4"
 
         # NVFP4 weights are repacked and written directly to gguf_writer.
@@ -1040,6 +1046,24 @@ class ModelBase:
 
         return part_names
 
+    @staticmethod
+    def load_hparams_guess(dir_model: Path) -> dict[str, Any] | None:
+        # some models ship no config.json, will try to guess them
+        from conversion import load_all_models
+        load_all_models()
+
+        for matcher, loader in ModelBase._hparams_loaders:
+            if matcher(dir_model):
+                return loader(dir_model)
+        return None
+
+    @classmethod
+    def register_hparams_loader(cls, matcher: HparamsMatcher) -> Callable[[HparamsLoader], HparamsLoader]:
+        def inner(loader: HparamsLoader) -> HparamsLoader:
+            cls._hparams_loaders.append((matcher, loader))
+            return loader
+        return inner
+
     @staticmethod
     def load_hparams(dir_model: Path, is_mistral_format: bool):
         if is_mistral_format:
@@ -1053,6 +1077,10 @@ class ModelBase:
             config = AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
         except Exception as e:
             logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            if not (dir_model / "config.json").is_file():
+                config = ModelBase.load_hparams_guess(dir_model)
+                if config is not None:
+                    return config
             logger.warning("Trying to load config.json instead")
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 config = json.load(f)

conversion/nemotron.py

@@ -275,10 +275,18 @@ class NemotronHModel(GraniteHybridModel):
                 return None
         elif cls.mtp_only:
             # --mtp: export the MTP head plus the tensors it shares with the target model
+            # Include lm_head scale sidecars so NVFP4 packing sees them.
             keep = name in (
                 "backbone.embeddings.weight",
                 "backbone.norm_f.weight",
                 "lm_head.weight",
+                "lm_head.weight_scale",
+                "lm_head.weight_scale_2",
+                "lm_head.weight_scale_inv",
+                "lm_head.input_scale",
+                "lm_head.input_global_scale",
+                "lm_head.weight_global_scale",
+                "lm_head.weight_packed",
             )
             if not keep:
                 return None

conversion/pockettts.py

@@ -0,0 +1,378 @@
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import Any, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, SentencePieceTokenTypes, TextModel, gguf, logger
+
+# Pocket TTS is a CALM: the backbone conditions a flow-matching decoder that generates one
+# continuous 32-d latent per frame. There is no codebook in this model.
+# The checkpoint ships no config.json, hparams come from _load_hparams() below.
+#
+# Tricks being used to support this model via existing llama.cpp code paths:
+# - bos_before_voice and bos_emb are learned input vectors, not tokens
+#   they are appended to the embedding table as extra tokens, to be looked up like any other row
+# - bos_emb lives in latent space, so input_linear is folded into it here
+# - the backbone has no lm_head, the embedding table is reused as output for the unused logits
+#
+# pipeline stage mapping:
+#   mimi encoder + speaker_proj --> mapped to normal mtmd audio encoder
+#   flow_lm.transformer         --> mapped to normal libllama text model (autoregressive)
+#   flow_lm.flow_net + out_eos  --> MTMD_GEN_PROCESS_TYPE_GEN_CODE
+#   mimi decoder                --> MTMD_GEN_PROCESS_TYPE_GEN_WAV
+
+# indices into mimi.encoder.model / mimi.decoder.model for stage i, see SEANetEncoder/SEANetDecoder
+_ENC_RES_IDX   = lambda i: 1 + 3 * i  # noqa: E731
+_ENC_SCALE_IDX = lambda i: 3 + 3 * i  # noqa: E731
+_DEC_SCALE_IDX = lambda i: 2 + 3 * i  # noqa: E731
+_DEC_RES_IDX   = lambda i: 3 + 3 * i  # noqa: E731
+
+_N_SEANET_STAGES = 3
+_SAMPLE_RATE = 24000
+
+
+def _tensor_shapes(dir_model: Path) -> dict[str, tuple[int, ...]]:
+    part_names = ModelBase.get_model_part_names(dir_model, "model", ".safetensors")
+    if len(part_names) != 1:
+        return {}
+    with gguf.utility.SafetensorsLocal(dir_model / part_names[0]) as part:
+        return {name: tuple(part[name].shape) for name in part.keys()}
+
+
+@ModelBase.register_hparams_loader(lambda dir_model: "flow_lm.bos_emb" in _tensor_shapes(dir_model))
+def _load_hparams(dir_model: Path) -> dict[str, Any]:
+    logger.info("gguf: detected pocket-tts checkpoint, deriving hparams from tensor shapes")
+    shapes = _tensor_shapes(dir_model)
+    n_vocab, n_embd = shapes["flow_lm.conditioner.embed.weight"]
+    n_layer = sum(1 for name in shapes if re.fullmatch(r"flow_lm\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_layer_a = sum(1 for name in shapes if re.fullmatch(r"mimi\.encoder_transformer\.transformer\.layers\.\d+\.norm1\.weight", name))
+    n_embd_a = shapes["mimi.encoder_transformer.transformer.layers.0.norm1.weight"][0]
+    return {
+        "architectures": ["PocketTTSModel"],
+        "model_type": "pockettts",
+        "num_hidden_layers": n_layer,
+        "hidden_size": n_embd,
+        "intermediate_size": shapes["flow_lm.transformer.layers.0.linear1.weight"][0],
+        # the transformer is fully causal with no context limit, this only bounds the KV cache
+        "max_position_embeddings": 4096,
+        # not in the checkpoint, but every released variant uses head_dim 64
+        "num_attention_heads": n_embd // 64,
+        # extra rows for the learned input vectors, see _embd_table()
+        "vocab_size": n_vocab + (2 if "flow_lm.bos_before_voice" in shapes else 1),
+        "rope_theta": 10000.0,
+        "layer_norm_eps": 1e-5,
+        "audio_config": {
+            "num_hidden_layers": n_layer_a,
+            "hidden_size": n_embd_a,
+            "intermediate_size": shapes["mimi.encoder_transformer.transformer.layers.0.linear1.weight"][0],
+            "num_attention_heads": n_embd_a // 64,
+        },
+    }
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.POCKETTTS
+
+    _LAYER_TENSOR_MAP = {
+        "norm1":               gguf.MODEL_TENSOR.ATTN_NORM,
+        "norm2":               gguf.MODEL_TENSOR.FFN_NORM,
+        "self_attn.out_proj":  gguf.MODEL_TENSOR.ATTN_OUT,
+        "linear1":             gguf.MODEL_TENSOR.FFN_UP,
+        "linear2":             gguf.MODEL_TENSOR.FFN_DOWN,
+    }
+
+    def set_vocab(self):
+        # this is a unigram sentencepiece model, llama.cpp's SPM tokenizer cannot do
+        # unigram segmentation, so use the UGM tokenizer instead
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        proto = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        proto.ParseFromString(open(self.dir_model / "tokenizer.model", "rb").read())
+        assert proto.trainer_spec.model_type == 1, "expected a unigram tokenizer"
+
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        # the last rows of the embedding table are not sentencepiece pieces
+        extra = self._extra_tokens()
+        for i, name in enumerate(extra):
+            tokens[len(tokens) - len(extra) + i] = name.encode("utf-8")
+            toktypes[len(tokens) - len(extra) + i] = SentencePieceTokenTypes.CONTROL
+            scores[len(tokens) - len(extra) + i] = -1000.0
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(proto.normalizer_spec.add_dummy_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(proto.normalizer_spec.remove_extra_whitespaces)
+        if proto.normalizer_spec.precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(proto.normalizer_spec.precompiled_charsmap)
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(False)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if not name.startswith("flow_lm."):
+            return  # mimi and the flow net go to the mmproj
+
+        if name == "flow_lm.conditioner.embed.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD), self._embd_table(data_torch))
+            return
+
+        if name.startswith("flow_lm.out_norm."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.transformer.layers."):
+            assert bid is not None
+            key_with_suffix = name.split(f"layers.{bid}.", 1)[1]
+            key, suffix = key_with_suffix.rsplit(".", 1)
+
+            if key == "self_attn.in_proj":
+                q, k, v = data_torch.chunk(3, dim=0)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k)
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+                return
+
+            tensor = self._LAYER_TENSOR_MAP.get(key)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix="." + suffix), data_torch)
+                return
+
+        return
+
+    def _extra_tokens(self) -> list[str]:
+        # the conditioner's padding row, then the learned vectors appended by _embd_table().
+        # bos_before_voice only exists when the pack sets insert_bos_before_voice
+        names = ["<|pad|>"]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            names.append("<|bos_before_voice|>")
+        names.append("<|audio_bos|>")
+        return names
+
+    def _embd_table(self, embed: Tensor) -> Tensor:
+        rows = [embed]
+        if "flow_lm.bos_before_voice" in self.model_tensors:
+            rows.append(self.model_tensors["flow_lm.bos_before_voice"]().reshape(1, -1).to(embed.dtype))
+
+        # bos_emb is a latent, it only enters the backbone through input_linear
+        bos_emb = self.model_tensors["flow_lm.bos_emb"]()
+        input_linear = self.model_tensors["flow_lm.input_linear.weight"]()
+        audio_bos = torch.nn.functional.linear(bos_emb.float(), input_linear.float()).reshape(1, -1)
+        rows.append(audio_bos.to(embed.dtype))
+
+        return torch.cat(rows, dim=0)
+
+
+@ModelBase.register("PocketTTSModel")
+class PocketTTSMmprojModel(MmprojModel):
+    has_audio_encoder = True
+    has_vision_encoder = False
+
+    _MIMI_TFM_MAP = {
+        "norm1":              (gguf.MODEL_TENSOR.A_ENC_INPUT_NORM,  gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_NORM),
+        "norm2":              (gguf.MODEL_TENSOR.A_ENC_OUTPUT_NORM, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_NORM),
+        "self_attn.out_proj": (gguf.MODEL_TENSOR.A_ENC_OUTPUT,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_OUT),
+        "linear1":            (gguf.MODEL_TENSOR.A_ENC_FFN_UP,      gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_UP),
+        "linear2":            (gguf.MODEL_TENSOR.A_ENC_FFN_DOWN,    gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_DOWN),
+        "layer_scale_1.scale": (gguf.MODEL_TENSOR.A_ENC_ATTN_SCALE, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_SCALE),
+        "layer_scale_2.scale": (gguf.MODEL_TENSOR.A_ENC_FFN_SCALE_LS, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_FFN_SCALE),
+    }
+    _MIMI_TFM_QKV = (
+        (gguf.MODEL_TENSOR.A_ENC_ATTN_Q, gguf.MODEL_TENSOR.A_ENC_ATTN_K, gguf.MODEL_TENSOR.A_ENC_ATTN_V),
+        (gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V),
+    )
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        assert self.hparams_audio is not None
+
+        # voice-prompt encoder: mimi encoder + speaker_proj
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        # note: the 24kHz sample rate is hardcoded on the clip.cpp side, like the other audio models
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_SPKENC)
+        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+        # mimi convolves the waveform directly, it is passed around as a 1-row "mel"
+        self.gguf_writer.add_audio_num_mel_bins(1)
+
+        # generation: flow-matching decoder + mimi decoder
+        # the SEANet and flow net hparams are constant across the family, clip.cpp holds them
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_GEN)
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(self.hparams_audio["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(self.hparams_audio["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(self.hparams_audio["num_attention_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(1e-5)
+
+        self.gguf_writer.add_gen_audio_model_variant(self.dir_model.name)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del name, bid, n_dims
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if ".seanet." in new_name or new_name in ("a.downsample.conv.weight", "a.gen.wav.upsample.weight"):
+            return gguf.GGMLQuantizationType.F16
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # the block index of the mimi transformers is parsed here, not by the base class
+        T = gguf.MODEL_TENSOR
+
+        if name in ("flow_lm.bos_emb", "flow_lm.bos_before_voice", "flow_lm.conditioner.embed.weight"):
+            return  # folded into the backbone embedding table
+        if name.startswith("flow_lm.transformer.") or name.startswith("flow_lm.out_norm."):
+            return  # backbone
+
+        if name == "flow_lm.speaker_proj_weight":
+            yield (self.format_tensor_name(T.A_ENC_SPEAKER_PROJ), data_torch)
+            return
+        if name == "flow_lm.input_linear.weight":
+            yield (self.format_tensor_name(T.A_GEN_INPUT_LINEAR), data_torch)
+            return
+        if name == "flow_lm.emb_mean":
+            yield (self.format_tensor_name(T.A_GEN_EMB_MEAN, suffix=""), data_torch)
+            return
+        if name == "flow_lm.emb_std":
+            yield (self.format_tensor_name(T.A_GEN_EMB_STD, suffix=""), data_torch)
+            return
+        if name.startswith("flow_lm.out_eos."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(T.A_GEN_OUT_EOS, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("flow_lm.flow_net."):
+            yield from self._flow_net_tensor(name, data_torch)
+            return
+
+        if name == "mimi.downsample.conv.conv.weight":
+            yield (self.format_tensor_name(T.A_ENC_DOWNSAMPLE_CONV), data_torch)
+            return
+        if name == "mimi.upsample.convtr.convtr.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_UPSAMPLE), data_torch)
+            return
+        if name == "mimi.quantizer.output_proj.weight":
+            yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_OUT), data_torch.squeeze(-1))
+            return
+
+        if "_transformer.transformer.layers." in name:
+            yield from self._mimi_tfm_tensor(name, data_torch)
+            return
+
+        if name.startswith("mimi.encoder.model.") or name.startswith("mimi.decoder.model."):
+            yield from self._seanet_tensor(name, data_torch)
+            return
+
+        return
+
+    def _flow_net_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        key = name.split("flow_lm.flow_net.", 1)[1]
+        suffix = "." + key.rsplit(".", 1)[1]
+
+        simple = {
+            "input_proj":                 T.A_GEN_FLOW_INPUT_PROJ,
+            "cond_embed":                 T.A_GEN_FLOW_COND_EMBD,
+            "final_layer.linear":         T.A_GEN_FLOW_FINAL_PROJ,
+            "final_layer.adaLN_modulation.1": T.A_GEN_FLOW_FINAL_ADA,
+        }
+        tensor = simple.get(key.rsplit(".", 1)[0])
+        if tensor is not None:
+            yield (self.format_tensor_name(tensor, suffix=suffix), data_torch)
+            return
+
+        if key.startswith("time_embed."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"time_embed.{bid}.", 1)[1]
+            time_map = {
+                "freqs":       (T.A_GEN_FLOW_TIME_FREQS, ""),
+                "mlp.0":       (T.A_GEN_FLOW_TIME_UP,    suffix),
+                "mlp.2":       (T.A_GEN_FLOW_TIME_DOWN,  suffix),
+                "mlp.3.alpha": (T.A_GEN_FLOW_TIME_NORM,  ""),
+            }
+            entry = time_map.get(rest) or time_map.get(rest.rsplit(".", 1)[0])
+            if entry is not None:
+                yield (self.format_tensor_name(entry[0], bid, suffix=entry[1]), data_torch)
+            return
+
+        if key.startswith("res_blocks."):
+            bid = int(key.split(".")[1])
+            rest = key.split(f"res_blocks.{bid}.", 1)[1].rsplit(".", 1)[0]
+            blk_map = {
+                "in_ln":                T.A_GEN_FLOW_BLK_NORM,
+                "mlp.0":                T.A_GEN_FLOW_BLK_UP,
+                "mlp.2":                T.A_GEN_FLOW_BLK_DOWN,
+                "adaLN_modulation.1":   T.A_GEN_FLOW_BLK_ADA,
+            }
+            tensor = blk_map.get(rest)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+            return
+
+    def _mimi_tfm_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        is_decoder = name.startswith("mimi.decoder_transformer.")
+        bid = int(name.split("_transformer.transformer.layers.", 1)[1].split(".")[0])
+        key_with_suffix = name.split(f".layers.{bid}.", 1)[1]
+
+        if key_with_suffix == "self_attn.in_proj.weight":
+            q, k, v = data_torch.chunk(3, dim=0)
+            names = self._MIMI_TFM_QKV[1 if is_decoder else 0]
+            for tensor, part in zip(names, (q, k, v)):
+                yield (self.format_tensor_name(tensor, bid), part)
+            return
+
+        key, suffix = key_with_suffix.rsplit(".", 1)
+        entry = self._MIMI_TFM_MAP.get(key) or self._MIMI_TFM_MAP.get(key_with_suffix)
+        if entry is None:
+            return
+        tensor = entry[1 if is_decoder else 0]
+        suffix = ".weight" if key_with_suffix.endswith(".scale") else "." + suffix
+        yield (self.format_tensor_name(tensor, bid, suffix=suffix), data_torch)
+
+    def _seanet_tensor(self, name: str, data_torch: Tensor) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        is_decoder = name.startswith("mimi.decoder.")
+        idx = int(name.split(".model.", 1)[1].split(".")[0])
+        suffix = "." + name.rsplit(".", 1)[1]
+
+        conv_in, conv_out, res1, res2, scale = (
+            (T.A_GEN_WAV_SEANET_CONV_IN, T.A_GEN_WAV_SEANET_CONV_OUT, T.A_GEN_WAV_SEANET_RES_CONV1,
+             T.A_GEN_WAV_SEANET_RES_CONV2, T.A_GEN_WAV_SEANET_SCALE_CONV)
+            if is_decoder else
+            (T.A_ENC_SEANET_CONV_IN, T.A_ENC_SEANET_CONV_OUT, T.A_ENC_SEANET_RES_CONV1,
+             T.A_ENC_SEANET_RES_CONV2, T.A_ENC_SEANET_SCALE_CONV)
+        )
+
+        if idx == 0:
+            yield (self.format_tensor_name(conv_in, suffix=suffix), data_torch)
+            return
+        if idx == 3 * _N_SEANET_STAGES + 2:
+            yield (self.format_tensor_name(conv_out, suffix=suffix), data_torch)
+            return
+
+        for stage in range(_N_SEANET_STAGES):
+            res_idx = _DEC_RES_IDX(stage) if is_decoder else _ENC_RES_IDX(stage)
+            scale_idx = _DEC_SCALE_IDX(stage) if is_decoder else _ENC_SCALE_IDX(stage)
+            if idx == scale_idx:
+                yield (self.format_tensor_name(scale, stage, suffix=suffix), data_torch)
+                return
+            if idx == res_idx:
+                # block.1 is the dilated conv, block.3 the pointwise one (0 and 2 are ELU)
+                inner = int(name.split(".block.", 1)[1].split(".")[0])
+                tensor = res1 if inner == 1 else res2
+                yield (self.format_tensor_name(tensor, stage, suffix=suffix), data_torch)
+                return

conversion/qwen.py

@@ -647,10 +647,13 @@ class DFlashModel(Qwen3Model):
         # own tokenizer logic, not the Qwen default).
         from . import get_model_class
         with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
-            target_arch = json.load(f)["architectures"][0]
+            target_hparams = json.load(f)
+            target_arch = target_hparams["architectures"][0]
         target_cls = get_model_class(target_arch)
 
         if target_cls is not type(self):
+            if target_arch == "NemotronHForCausalLM":
+                setattr(self, "is_moe", "num_experts_per_tok" in target_hparams)
             target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
         else:
             super().set_vocab()
@@ -688,6 +691,12 @@ class DFlashModel(Qwen3Model):
             name = "model." + name
         return super().filter_tensors((name, gen))
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "model.embed_tokens.weight" and not self.hparams.get("has_embed_tokens", True):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
 
 @ModelBase.register("Qwen3DSparkModel")
 class DSparkModel(DFlashModel):

examples/model-conversion/requirements.txt

@@ -1,6 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch
-torchvision
+torchvision; platform_machine != "s390x"
 transformers
 huggingface-hub
 accelerate

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.py

@@ -2,12 +2,15 @@
 
 import argparse
 import os
+import sys
 import importlib
 import torch
 import numpy as np
 
 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM
-from pathlib import Path
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+from utils.common import save_output_data
 
 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 
@@ -54,6 +57,7 @@ print(f"Model name: {model_name}")
 
 prompt = "Hello world today"
 input_ids = tokenizer(prompt, return_tensors="pt").input_ids  # ty: ignore[call-non-callable]
+token_ids = input_ids[0].cpu().tolist()
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")
 print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")  # ty: ignore[unresolved-attribute]
@@ -74,21 +78,8 @@ with torch.no_grad():
     print(f"Hidden dimension: {token_embeddings.shape[-1]}")
     print(f"Number of tokens: {token_embeddings.shape[0]}")
 
-    # Save raw token embeddings
-    data_dir = Path("data")
-    data_dir.mkdir(exist_ok=True)
-    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
-    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
-
-    # Save all token embeddings as binary
     print(token_embeddings)
-    token_embeddings.astype(np.float32).tofile(bin_filename)
-
-    # Save as text for inspection
-    with open(txt_filename, "w") as f:
-        for i, embedding in enumerate(token_embeddings):
-            for j, val in enumerate(embedding):
-                f.write(f"{i} {j} {val:.6f}\n")
+    save_output_data(token_embeddings, token_ids, prompt, model_name, type_suffix="-embeddings")
 
     # Print embeddings per token in the requested format
     print("\nToken embeddings:")
@@ -110,5 +101,3 @@ with torch.no_grad():
     for i, token in enumerate(tokens):
         print(f"  Token {i}: {repr(token)}")
 
-    print(f"Saved bin logits to: {bin_filename}")
-    print(f"Saved txt logist to: {txt_filename}")

examples/speculative-simple/README.md

@@ -3,10 +3,47 @@
 Demonstration of basic greedy speculative decoding
 
 ```bash
+# spec-type draft-simple
 ./bin/llama-speculative-simple \
-    -m  ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
-    -md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
-    -f test.txt -c 0 -ngl 99 --color on \
-    --sampling-seq k --top-k 1 -fa on --temp 0.0 \
-    -ngld 99 --spec-draft-n-max 16 --spec-draft-n-draft-min 5 --draft-p-min 0.9
+        -hf  ggml-org/Qwen3-8B-Base-GGUF:Q8_0 \
+        -hfd ggml-org/Qwen3-0.6B-Base-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-simple --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3.6-27B-GGUF:Q8_0 \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-mtp (with shared KV cache)
+# note: this model needs a <s> token at the start to somewhat work without the chat template
+./bin/llama-speculative-simple \
+        -hf ggml-org/Gemma-4-31B-it-GGUF:Q8_0 \
+        -p "<s>Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-mtp --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-eagle3
+./bin/llama-speculative-simple \
+        -hf ggml-org/gpt-oss-20b-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-eagle3 --spec-draft-n-max 3 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dflash
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dflash --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
+
+# spec-type draft-dspark
+./bin/llama-speculative-simple \
+        -hf ggml-org/Qwen3-8B-GGUF \
+        -p "Here is a quick sort implementation in C++. Just code, no comments:\n\n#include" \
+        --spec-type draft-dspark --spec-draft-n-max 7 -ngld 99 --color on \
+        -n 256 --temp 0 --top-k 1 --seed 42 -ngl 99 -lv 4
 ```

examples/speculative-simple/speculative-simple.cpp

@@ -51,48 +51,23 @@ int main(int argc, char ** argv) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
 
-    // load the draft model
-    llama_model_ptr model_dft;
-    llama_context_ptr ctx_dft;
+    // load the draft model (if any) - this also creates the MTP draft context when MTP speculation is enabled
+    common_speculative_init_result_ptr spec_init;
 
-    // TODO: simplify this logic
     {
-        const auto & params_spec = params.speculative.draft;
+        common_params params_dft = common_base_params_to_speculative(params);
 
-        auto params_dft = params;
-
-        params_dft.n_outputs_max = params.n_parallel;
-        params_dft.n_outputs_max_per_seq = 1;
-
-        params_dft.devices      = params_spec.devices;
-        params_dft.model        = params_spec.mparams;
-        params_dft.n_gpu_layers = params_spec.n_gpu_layers;
-
-        if (params_spec.cpuparams.n_threads > 0) {
-            params_dft.cpuparams.n_threads       = params.speculative.draft.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
-        }
-
-        params_dft.tensor_buft_overrides = params.speculative.draft.tensor_buft_overrides;
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-
-        model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
-        if (model_dft == nullptr) {
-            LOG_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
-            return 1;
-        }
-
-        auto cparams = common_context_params_to_llama(params_dft);
-        ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
+        spec_init = common_speculative_init_from_params(params_dft, model_tgt, ctx_tgt);
 
         params.speculative.draft.ctx_tgt = ctx_tgt;
-        params.speculative.draft.ctx_dft = ctx_dft.get();
+        params.speculative.draft.ctx_dft = spec_init->context();
     }
 
+    llama_context * ctx_dft = params.speculative.draft.ctx_dft;
+
     // check if the context supports partial sequence removal
-    const bool use_ckpt_tgt = (common_context_can_seq_rm(ctx_tgt)       == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
-    const bool use_ckpt_dft = (common_context_can_seq_rm(ctx_dft.get()) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
+    const bool use_ckpt_tgt = common_context_can_seq_rm(ctx_tgt) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+    const bool use_ckpt_dft = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
 
     if (use_ckpt_tgt) {
         LOG_INF("speculative decoding will use checkpoints (context does not support partial sequence removal)\n");
@@ -138,9 +113,30 @@ int main(int argc, char ** argv) {
     // target model sampling context
     common_sampler_ptr smpl(common_sampler_init(model_tgt, params.sampling));
 
-    // eval the prompt
-    llama_decode(ctx_tgt,       llama_batch_get_one(inp.data(), inp.size() - 1));
-    llama_decode(ctx_dft.get(), llama_batch_get_one(inp.data(), inp.size() - 1));
+    // init the speculator
+    const auto & params_spec = params.speculative;
+
+    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
+
+    if (spec == nullptr) {
+        LOG_ERR("%s", "failed to initialize speculative decoding\n");
+        return 1;
+    }
+
+    // eval the prompt on the target and feed it to the speculative implementation(s)
+    {
+        llama_batch batch_prompt = llama_batch_init(inp.size(), 0, 1);
+        for (size_t i = 0; i < inp.size() - 1; ++i) {
+            common_batch_add(batch_prompt, inp[i], i, { seq_id }, false);
+        }
+
+        llama_decode(ctx_tgt, batch_prompt);
+
+        if (!common_speculative_process(spec, batch_prompt)) {
+            LOG_ERR("%s", "failed to process speculative prompt\n");
+            return 1;
+        }
+    }
 
     // note: keep the last token separate!
     llama_token id_last = inp.back();
@@ -151,18 +147,12 @@ int main(int argc, char ** argv) {
 
     int n_past = inp.size() - 1;
 
-    // init the speculator
-    const auto & params_spec = params.speculative;
-
-    struct common_speculative * spec = common_speculative_init(params.speculative, 1);
-
     common_speculative_begin(spec, seq_id, prompt_tgt);
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 
-    size_t n_draft = 0;
-
     llama_tokens draft;
+
     common_prompt_checkpoint ckpt;
 
     const auto t_enc_end = ggml_time_us();
@@ -184,13 +174,20 @@ int main(int argc, char ** argv) {
                     llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), seq_id));
 
             if (use_ckpt_dft) {
-                ckpt.update_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                ckpt.update_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
             }
 
+            // determine the max draft that fits the remaining context and generation budget
+            int n_draft_max = (int) llama_n_ctx(ctx_tgt) - n_past - 2;
+            if (params.n_predict >= 0) {
+                n_draft_max = std::min(n_draft_max, params.n_predict - n_predict - 1);
+            }
+            n_draft_max = std::max(n_draft_max, 0);
+
             // generate a new draft
             common_speculative_get_draft_params(spec, seq_id) = {
                 /* .drafting   = */ true,
-                /* .n_max      = */ -1,
+                /* .n_max      = */ n_draft_max,
                 /* .n_past     = */ n_past,
                 /* .id_last    = */ id_last,
                 /* .prompt     = */ &prompt_tgt,
@@ -198,9 +195,6 @@ int main(int argc, char ** argv) {
             };
             common_speculative_draft(spec);
 
-            // save the original draft size
-            n_draft = draft.size();
-
             // save a checkpoint of the target context before evaluating the draft
             // this allows us to restore the state if partial draft acceptance occurs
             if (!draft.empty()) {
@@ -209,10 +203,13 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            // reset the draft context to the checkpoint before verification
+            if (ctx_dft) {
+                if (use_ckpt_dft) {
+                    ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                }
 
-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
             }
         } else {
             // we have a previous (partial) draft to reuse from checkpoint restoration
@@ -236,10 +233,10 @@ int main(int argc, char ** argv) {
             llama_decode(ctx_tgt, batch_tgt);
         }
 
-        // evaluate the same batch with the draft model
-        {
-            // TODO: extend to support MTP, Eagle, etc. See server code for reference
-            llama_decode(ctx_dft.get(), batch_tgt);
+        // feed the batch to the speculative implementation(s) - this drives the draft model, MTP, Eagle3, etc.
+        if (!common_speculative_process(spec, batch_tgt)) {
+            LOG_ERR("%s", "failed to process speculative batch\n");
+            break;
         }
 
         // only save the sampler sampler state if we use checkpoints
@@ -248,6 +245,9 @@ int main(int argc, char ** argv) {
             smpl_save.reset(common_sampler_clone(smpl.get()));
         }
 
+        // save the size of the draft being verified
+        const size_t n_draft = draft.size();
+
         // sample from the full target batch and return the accepted tokens based on the target sampler
         //
         // for each token to be accepted, the sampler would have to sample that same token
@@ -264,8 +264,8 @@ int main(int argc, char ** argv) {
         // check for partial draft acceptance:
         // if the context doesn't support partial sequence removal, restore the checkpoint
         // and make the accepted tokens the new partial draft for the next iteration
-        if (use_ckpt_tgt && ids.size() - 1 < draft.size()) {
-            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, draft.size());
+        if (use_ckpt_tgt && ids.size() - 1 < n_draft) {
+            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, n_draft);
 
             draft = std::move(ids);
 
@@ -275,10 +275,10 @@ int main(int argc, char ** argv) {
                 llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, ckpt.pos_max + 1, -1);
             }
 
-            {
-                ckpt.load_dft(ctx_dft.get(), seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            if (ctx_dft) {
+                ckpt.load_dft(ctx_dft, seq_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
 
-                llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, ckpt.pos_max + 1, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, ckpt.pos_max + 1, -1);
             }
 
             prompt_tgt.resize(ckpt.n_tokens);
@@ -329,8 +329,11 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_memory_seq_rm(llama_get_memory(ctx_tgt),       seq_id, n_past, -1);
-            llama_memory_seq_rm(llama_get_memory(ctx_dft.get()), seq_id, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), seq_id, n_past, -1);
+
+            if (ctx_dft) {
+                llama_memory_seq_rm(llama_get_memory(ctx_dft), seq_id, n_past, -1);
+            }
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
@@ -356,6 +359,7 @@ int main(int argc, char ** argv) {
 
     LOG_INF("\n");
     LOG_INF("draft:\n\n");
+    common_speculative_print_stats(spec);
 
     LOG_INF("\n");
     LOG_INF("target:\n\n");

ggml/include/ggml-backend.h

@@ -154,6 +154,8 @@ extern "C" {
         bool buffer_from_host_ptr;
         // event synchronization
         bool events;
+        // mmap is supported for loading
+        bool mmap_support;
     };
 
     // all the device properties

ggml/src/ggml-backend-meta.cpp

@@ -132,6 +132,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ false, // Not implemented.
         /* .buffer_from_host_ptr  = */ false, // Not implemented.
         /* .events                = */ false, // Not implemented.
+        /* .mmap_support          = */ true,
     };
     for (ggml_backend_dev_t simple_dev : meta_dev_ctx->simple_devs) {
         ggml_backend_dev_props tmp_props;
@@ -140,6 +141,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
         props->caps.host_buffer          = props->caps.host_buffer          && tmp_props.caps.host_buffer;
         props->caps.buffer_from_host_ptr = props->caps.buffer_from_host_ptr && tmp_props.caps.buffer_from_host_ptr;
         props->caps.events               = props->caps.events               && tmp_props.caps.events;
+        props->caps.mmap_support         = props->caps.mmap_support         && tmp_props.caps.mmap_support;
     }
 }
 

ggml/src/ggml-blas/ggml-blas.cpp

@@ -367,6 +367,7 @@ static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct gg
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ true,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2815,6 +2815,7 @@ static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ true,
+        /* .mmap_support          = */ true,
     };
 }
 

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -397,6 +397,7 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ true,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

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

@@ -1865,6 +1865,37 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     ggml_cuda_mul_mat_cublas(ctx, src0, src1, dst);
 }
 
+// returns true when ggml_cuda_mul_mat_id takes the fallback path that requires stream synchronization
+// [TAG_MUL_MAT_ID_CUDA_GRAPHS]
+static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int cc) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return true;
+    }
+
+    if (dst->ne[2] <= MMVQ_MAX_BATCH_SIZE) {
+        if (ggml_is_quantized(src0->type)) {
+            if (dst->ne[2] <= get_mmvq_mmid_max_batch(src0->type, cc)) {
+                return false;
+            }
+        } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+            return false;
+        }
+    }
+
+    if (ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[2], /*n_experts=*/src0->ne[2])) {
+        return false;
+    }
+
+    if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
+        return false;
+    }
+
+    return true;
+}
+
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
@@ -1907,7 +1938,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
     }
 
     // note: this path should not be reached when recording CUDA graphs, because it requires stream synchronization
-    // TODO: add asserts to verify this. should work with CUDA, HIP, etc.
+    GGML_ASSERT(ggml_cuda_mul_mat_id_needs_sync(dst, cc));
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(nb12 % nb11 == 0);
@@ -2522,10 +2553,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
         // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
         if (node->op == GGML_OP_MUL_MAT_ID) {
             const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
-            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
-                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+            if (ggml_cuda_mul_mat_id_needs_sync(node, cc)) {
+                // the mul_mat_id fallback path synchronizes the stream, so we cannot use CUDA graphs
                 // ref: https://github.com/ggml-org/llama.cpp/pull/18958
                 use_cuda_graph = false;
 #ifndef NDEBUG
@@ -4801,6 +4830,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ events,
+        /* .mmap_support          = */ props->type != GGML_BACKEND_DEVICE_TYPE_IGPU,
     };
 }
 

ggml/src/ggml-cuda/wkv.cu

@@ -141,6 +141,57 @@ static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, cons
     }
 }
 
+template <int rows_per_block>
+static __global__ void __launch_bounds__(WARP_SIZE * rows_per_block, 2)
+rwkv_wkv7_f32_t1_warp_row(const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
+    constexpr int head_size = CUDA_WKV_BLOCK_SIZE;
+    constexpr int half_head = head_size / 2;
+
+    const int lane = threadIdx.x;
+    const int row  = blockIdx.y * rows_per_block + threadIdx.y;
+    const int bid  = blockIdx.x;
+
+    const int batch_i = bid / H;
+    const int head_i  = bid % H;
+    const int state_size = C * head_size;
+    const int head_off = head_i * head_size;
+    const int t = batch_i * C + head_off + row;
+
+    __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
+
+    if (threadIdx.y == 0) {
+        _r[lane] = r[batch_i * C + head_off + lane];
+        _w[lane] = w[batch_i * C + head_off + lane];
+        _k[lane] = k[batch_i * C + head_off + lane];
+        _a[lane] = a[batch_i * C + head_off + lane];
+        _b[lane] = b[batch_i * C + head_off + lane];
+
+        _r[lane + half_head] = r[batch_i * C + head_off + lane + half_head];
+        _w[lane + half_head] = w[batch_i * C + head_off + lane + half_head];
+        _k[lane + half_head] = k[batch_i * C + head_off + lane + half_head];
+        _a[lane + half_head] = a[batch_i * C + head_off + lane + half_head];
+        _b[lane + half_head] = b[batch_i * C + head_off + lane + half_head];
+    }
+    __syncthreads();
+
+    const int64_t state_base = batch_i * state_size + head_i * head_size * head_size + row * head_size;
+    const float s0 = s[state_base + lane];
+    const float s1 = s[state_base + lane + half_head];
+    const float sa = warp_reduce_sum(_a[lane] * s0 + _a[lane + half_head] * s1);
+
+    const float vt  = v[t];
+    const float st0 = s0 * _w[lane]             + _k[lane]             * vt + sa * _b[lane];
+    const float st1 = s1 * _w[lane + half_head] + _k[lane + half_head] * vt + sa * _b[lane + half_head];
+    const float y   = warp_reduce_sum(st0 * _r[lane] + st1 * _r[lane + half_head]);
+
+    dst[T * C + state_base + lane]             = st0;
+    dst[T * C + state_base + lane + half_head] = st1;
+
+    if (lane == 0) {
+        dst[t] = y;
+    }
+}
+
 void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float * k_d  = (const float *)dst->src[0]->data;
     const float * v_d  = (const float *)dst->src[1]->data;
@@ -191,7 +242,10 @@ void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     GGML_ASSERT(C % H == 0);
     GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
 
-    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+    if (T / B == 1 && C / H == CUDA_WKV_BLOCK_SIZE) {
+        constexpr int rows_per_block = 4;
+        rwkv_wkv7_f32_t1_warp_row<rows_per_block><<<dim3(B * H, CUDA_WKV_BLOCK_SIZE / rows_per_block), dim3(WARP_SIZE, rows_per_block), 0, stream>>>(T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    } else if (C / H == CUDA_WKV_BLOCK_SIZE) {
         rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
     } else {
         rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);

ggml/src/ggml-et/ggml-et.cpp

@@ -1646,6 +1646,7 @@ static void ggml_backend_et_device_get_props(ggml_backend_dev_t dev, struct ggml
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

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

@@ -3930,6 +3930,7 @@ static void ggml_backend_hexagon_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ (bool) opt_hostbuf,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ false,
     };
 }
 

ggml/src/ggml-metal/ggml-metal.cpp

@@ -681,6 +681,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ true,
         /* .events               = */ true,
+        /* .mmap_support         = */ true,
     };
 }
 

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -10795,6 +10795,7 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ false,
     };
 }
 

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -763,6 +763,7 @@ static void ggml_backend_openvino_device_get_props(ggml_backend_dev_t dev, ggml_
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1881,6 +1881,7 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -5649,6 +5649,7 @@ static void ggml_backend_sycl_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ events,
+        /* .mmap_support          = */ true,
     };
 }
 

ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp

@@ -111,6 +111,7 @@ uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_
     apir_encode_bool_t(enc, &props.caps.host_buffer);
     apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr);
     apir_encode_bool_t(enc, &props.caps.events);
+    apir_encode_bool_t(enc, &props.caps.mmap_support);
 
     return 0;
 }

ggml/src/ggml-virtgpu/backend/shared/api_remoting.h

@@ -7,7 +7,7 @@
 #include <cstdint>
 
 #define APIR_PROTOCOL_MAJOR 0
-#define APIR_PROTOCOL_MINOR 1
+#define APIR_PROTOCOL_MINOR 2
 
 #define APIR_HANDSHAKE_MAGIC 0xab1e
 

ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp

@@ -11,9 +11,9 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
 
     context->gpu = gpu;
 
-    bool async__unused, host_buffer__unused, events__unused;
+    bool async__unused, host_buffer__unused, events__unused, mmap_support__unused;
     bool buffer_from_host_ptr;
-    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused);
+    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused, &mmap_support__unused);
 
     if (buffer_from_host_ptr) {
         context->apir_context = apir_device_buffer_from_ptr(gpu, size, size);

ggml/src/ggml-virtgpu/ggml-backend-device.cpp

@@ -65,7 +65,7 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
 
     virtgpu * gpu = DEV_TO_GPU(dev);
     apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr,
-                          &props->caps.events);
+                          &props->caps.events, &props->caps.mmap_support);
 
     props->caps.buffer_from_host_ptr = false;
     props->caps.async                = false;

ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp

@@ -144,7 +144,8 @@ void apir_device_get_props(virtgpu * gpu,
                            bool *    async,
                            bool *    host_buffer,
                            bool *    buffer_from_host_ptr,
-                           bool *    events) {
+                           bool *    events,
+                           bool *    mmap_support) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -157,6 +158,7 @@ void apir_device_get_props(virtgpu * gpu,
     apir_decode_bool_t(decoder, host_buffer);
     apir_decode_bool_t(decoder, buffer_from_host_ptr);
     apir_decode_bool_t(decoder, events);
+    apir_decode_bool_t(decoder, mmap_support);
 
     remote_call_finish(gpu, encoder, decoder);
 

ggml/src/ggml-virtgpu/virtgpu-forward.gen.h

@@ -13,7 +13,8 @@ void                           apir_device_get_props(struct virtgpu * gpu,
                                                      bool *           async,
                                                      bool *           host_buffer,
                                                      bool *           buffer_from_host_ptr,
-                                                     bool *           events);
+                                                     bool *           events,
+                                                     bool *           mmap_support);
 apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
 
 /* buffer-type */

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -17891,6 +17891,7 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
         /* .host_buffer           = */ true,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ true,
+        /* .mmap_support          = */ !ctx->is_integrated_gpu,
     };
 }
 

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -2815,11 +2815,25 @@ class ggml_webgpu_shader_lib {
         key.common.v_direct &= decisions.use_sg_matrix && key.common.v_type == GGML_TYPE_F16;
         key.use_sg_matrix = decisions.use_sg_matrix;
 
-        const uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
+        uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
             context.wg_mem_limit_bytes, decisions.q_tile, decisions.use_sg_matrix ? context.sg_mat_n : 1u,
             key.common.head_dim_qk, key.common.head_dim_v, key.common.has_mask,
             key.common.k_direct || key.common.v_direct);
-        GGML_ASSERT(max_kv_tile > 0);
+
+        // WorkGroup storage size isn't enough for some params with subgroup matrices path (ref. https://github.com/ggml-org/llama.cpp/pull/26566)
+        if (max_kv_tile == 0) {
+            GGML_ASSERT(decisions.use_sg_matrix);
+            // switch to flash_attn_reg_tile path
+            decisions.use_sg_matrix = false;
+            decisions.q_tile        = GGML_WEBGPU_FLASH_ATTN_TILE_Q_TILE;
+            key.common.k_direct     = false;
+            key.common.v_direct     = false;
+            key.use_sg_matrix       = false;
+            max_kv_tile             = ggml_webgpu_flash_attn_max_kv_tile(
+                context.wg_mem_limit_bytes, decisions.q_tile, 1u, key.common.head_dim_qk, key.common.head_dim_v,
+                key.common.has_mask, key.common.k_direct || key.common.v_direct);
+            GGML_ASSERT(max_kv_tile > 0);
+        }
 
         decisions.kv_tile = decisions.use_sg_matrix ?
                                 std::min(max_kv_tile, context.sg_mat_n * GGML_WEBGPU_FLASH_ATTN_PREFERRED_KV_SG_TILES) :
@@ -2993,6 +3007,10 @@ class ggml_webgpu_shader_lib {
                 defines.push_back("SRC_F16");
                 variant += "_f16";
                 break;
+            case GGML_TYPE_I32:
+                defines.push_back("SRC_I32");
+                variant += "_i32";
+                break;
             default:
                 GGML_ABORT("Unsupported src type for cpy shader");
         }

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -3942,6 +3942,7 @@ static void ggml_backend_webgpu_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 
@@ -4283,9 +4284,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
             break;
         case GGML_OP_CPY:
         case GGML_OP_CONT:
-            supports_op = ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
-                           (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) ||
-                          (op->type == GGML_TYPE_I32 && src0->type == GGML_TYPE_F32);
+            supports_op = (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32) &&
+                          (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
             break;
         case GGML_OP_SET:
             supports_op = src0->type == src1->type && src0->type == op->type &&

ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl

@@ -4,6 +4,8 @@ enable f16;
 #define SRC_TYPE f32
 #elif defined(SRC_F16)
 #define SRC_TYPE f16
+#elif defined(SRC_I32)
+#define SRC_TYPE i32
 #endif
 
 #ifdef DST_F32

ggml/src/ggml-zdnn/ggml-zdnn.cpp

@@ -487,7 +487,8 @@ static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .async                = */ false,
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ false,
-        /* .events               = */ false
+        /* .events               = */ false,
+        /* .mmap_support         = */ true,
     };
 }
 

ggml/src/ggml-zendnn/ggml-zendnn.cpp

@@ -654,7 +654,8 @@ static void ggml_backend_zendnn_device_get_props(ggml_backend_dev_t dev, struct
         /* .async                = */ false,
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ true,
-        /* .events               = */ false
+        /* .events               = */ false,
+        /* .mmap_support         = */ true,
     };
 }
 

gguf-py/gguf/constants.py

@@ -407,6 +407,8 @@ class Keys:
 
     class ClipGenAudio:
         PROJECTOR_TYPE      = "clip.gen.audio.projector_type" # for mixed modality models
+        # name of the weight variant, for settings that are not in the checkpoint
+        MODEL_VARIANT       = "clip.gen.audio.model_variant"
         EMBEDDING_LENGTH    = "clip.gen.audio.embedding_length"
         FEED_FORWARD_LENGTH = "clip.gen.audio.feed_forward_length"
         BLOCK_COUNT         = "clip.gen.audio.block_count"
@@ -581,6 +583,7 @@ class MODEL_ARCH(IntEnum):
     MELLUM           = auto()
     NANBEIGE         = auto()
     QWEN3TTS         = auto()
+    POCKETTTS        = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -1040,6 +1043,38 @@ class MODEL_TENSOR(IntEnum):
     A_GEN_WAV_DAC_RES_CONV2        = auto() # DAC residual unit, pointwise causal conv
     A_GEN_WAV_DAC_POST_SNAKE       = auto() # DAC final SnakeBeta
     A_GEN_WAV_DAC_POST_CONV        = auto() # DAC conv_post -> 1-channel PCM
+    # pocket-tts: SEANet encoder (speaker path) and decoder (a.gen.wav path)
+    A_ENC_SEANET_CONV_IN     = auto()
+    A_ENC_SEANET_CONV_OUT    = auto()
+    A_ENC_SEANET_RES_CONV1   = auto() # residual unit, dilated conv
+    A_ENC_SEANET_RES_CONV2   = auto() # residual unit, pointwise conv
+    A_ENC_SEANET_SCALE_CONV  = auto() # strided downsample conv
+    A_ENC_ATTN_SCALE         = auto() # layer scale (gamma) on the attn output
+    A_ENC_FFN_SCALE_LS       = auto() # layer scale (gamma) on the FFN output
+    A_ENC_SPEAKER_PROJ       = auto() # voice latent -> backbone embd
+    A_GEN_FLOW_INPUT_PROJ    = auto()
+    A_GEN_FLOW_COND_EMBD     = auto()
+    A_GEN_FLOW_TIME_FREQS    = auto() # timestep embedder, stored cos/sin frequencies
+    A_GEN_FLOW_TIME_UP       = auto()
+    A_GEN_FLOW_TIME_DOWN     = auto()
+    A_GEN_FLOW_TIME_NORM     = auto() # RMSNorm alpha
+    A_GEN_FLOW_BLK_NORM      = auto() # AdaLN res block, in_ln
+    A_GEN_FLOW_BLK_UP        = auto()
+    A_GEN_FLOW_BLK_DOWN      = auto()
+    A_GEN_FLOW_BLK_ADA       = auto() # AdaLN modulation, -> shift/scale/gate
+    A_GEN_FLOW_FINAL_ADA     = auto() # final layer AdaLN modulation, -> shift/scale
+    A_GEN_FLOW_FINAL_PROJ    = auto()
+    A_GEN_OUT_EOS            = auto() # end-of-speech head on the backbone hidden state
+    A_GEN_INPUT_LINEAR       = auto() # generated latent -> backbone embd
+    A_GEN_EMB_MEAN           = auto() # latent denormalization stats
+    A_GEN_EMB_STD            = auto()
+    A_GEN_WAV_QUANT_OUT      = auto() # DummyQuantizer output_proj, latent -> decoder dim
+    A_GEN_WAV_UPSAMPLE       = auto() # frame rate -> encoder frame rate, depthwise convtr
+    A_GEN_WAV_SEANET_CONV_IN   = auto()
+    A_GEN_WAV_SEANET_CONV_OUT  = auto() # -> 1-channel PCM
+    A_GEN_WAV_SEANET_RES_CONV1 = auto()
+    A_GEN_WAV_SEANET_RES_CONV2 = auto()
+    A_GEN_WAV_SEANET_SCALE_CONV = auto() # strided upsample convtr
     A_MMPROJ              = auto()
     A_MMPROJ_FC           = auto()
     A_MM_NORM_PRE         = auto()
@@ -1255,6 +1290,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.MELLUM:           "mellum",
     MODEL_ARCH.NANBEIGE:         "nanbeige",
     MODEL_ARCH.QWEN3TTS:         "qwen3tts",
+    MODEL_ARCH.POCKETTTS:        "pockettts",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1709,6 +1745,37 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV2:   "a.gen.wav.dac.blk.{bid}.res.{xid}.conv2",
     MODEL_TENSOR.A_GEN_WAV_DAC_POST_SNAKE:  "a.gen.wav.dac.post_snake",
     MODEL_TENSOR.A_GEN_WAV_DAC_POST_CONV:   "a.gen.wav.dac.post_conv",
+    MODEL_TENSOR.A_ENC_SEANET_CONV_IN:      "a.seanet.conv_in",
+    MODEL_TENSOR.A_ENC_SEANET_CONV_OUT:     "a.seanet.conv_out",
+    MODEL_TENSOR.A_ENC_SEANET_RES_CONV1:    "a.seanet.blk.{bid}.res_conv1",
+    MODEL_TENSOR.A_ENC_SEANET_RES_CONV2:    "a.seanet.blk.{bid}.res_conv2",
+    MODEL_TENSOR.A_ENC_SEANET_SCALE_CONV:   "a.seanet.blk.{bid}.scale_conv",
+    MODEL_TENSOR.A_ENC_ATTN_SCALE:          "a.blk.{bid}.ls1",
+    MODEL_TENSOR.A_ENC_FFN_SCALE_LS:        "a.blk.{bid}.ls2",
+    MODEL_TENSOR.A_ENC_SPEAKER_PROJ:        "a.speaker_proj",
+    MODEL_TENSOR.A_GEN_FLOW_INPUT_PROJ:     "a.gen.flow.input_proj",
+    MODEL_TENSOR.A_GEN_FLOW_COND_EMBD:      "a.gen.flow.cond_embd",
+    MODEL_TENSOR.A_GEN_FLOW_TIME_FREQS:     "a.gen.flow.time.{bid}.freqs",
+    MODEL_TENSOR.A_GEN_FLOW_TIME_UP:        "a.gen.flow.time.{bid}.up",
+    MODEL_TENSOR.A_GEN_FLOW_TIME_DOWN:      "a.gen.flow.time.{bid}.down",
+    MODEL_TENSOR.A_GEN_FLOW_TIME_NORM:      "a.gen.flow.time.{bid}.norm",
+    MODEL_TENSOR.A_GEN_FLOW_BLK_NORM:       "a.gen.flow.blk.{bid}.norm",
+    MODEL_TENSOR.A_GEN_FLOW_BLK_UP:         "a.gen.flow.blk.{bid}.up",
+    MODEL_TENSOR.A_GEN_FLOW_BLK_DOWN:       "a.gen.flow.blk.{bid}.down",
+    MODEL_TENSOR.A_GEN_FLOW_BLK_ADA:        "a.gen.flow.blk.{bid}.ada",
+    MODEL_TENSOR.A_GEN_FLOW_FINAL_ADA:      "a.gen.flow.final.ada",
+    MODEL_TENSOR.A_GEN_FLOW_FINAL_PROJ:     "a.gen.flow.final.proj",
+    MODEL_TENSOR.A_GEN_OUT_EOS:             "a.gen.out_eos",
+    MODEL_TENSOR.A_GEN_INPUT_LINEAR:        "a.gen.input_linear",
+    MODEL_TENSOR.A_GEN_EMB_MEAN:            "a.gen.emb_mean",
+    MODEL_TENSOR.A_GEN_EMB_STD:             "a.gen.emb_std",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_OUT:       "a.gen.wav.quant_out",
+    MODEL_TENSOR.A_GEN_WAV_UPSAMPLE:        "a.gen.wav.upsample",
+    MODEL_TENSOR.A_GEN_WAV_SEANET_CONV_IN:  "a.gen.wav.seanet.conv_in",
+    MODEL_TENSOR.A_GEN_WAV_SEANET_CONV_OUT: "a.gen.wav.seanet.conv_out",
+    MODEL_TENSOR.A_GEN_WAV_SEANET_RES_CONV1: "a.gen.wav.seanet.blk.{bid}.res_conv1",
+    MODEL_TENSOR.A_GEN_WAV_SEANET_RES_CONV2: "a.gen.wav.seanet.blk.{bid}.res_conv2",
+    MODEL_TENSOR.A_GEN_WAV_SEANET_SCALE_CONV: "a.gen.wav.seanet.blk.{bid}.scale_conv",
     MODEL_TENSOR.A_MMPROJ:                  "mm.a.mlp.{bid}",
     MODEL_TENSOR.A_MMPROJ_FC:               "mm.a.fc",
     MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
@@ -2020,6 +2087,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV2,
         MODEL_TENSOR.A_GEN_WAV_DAC_POST_SNAKE,
         MODEL_TENSOR.A_GEN_WAV_DAC_POST_CONV,
+        MODEL_TENSOR.A_ENC_SEANET_CONV_IN,
+        MODEL_TENSOR.A_ENC_SEANET_CONV_OUT,
+        MODEL_TENSOR.A_ENC_SEANET_RES_CONV1,
+        MODEL_TENSOR.A_ENC_SEANET_RES_CONV2,
+        MODEL_TENSOR.A_ENC_SEANET_SCALE_CONV,
+        MODEL_TENSOR.A_ENC_ATTN_SCALE,
+        MODEL_TENSOR.A_ENC_FFN_SCALE_LS,
+        MODEL_TENSOR.A_ENC_SPEAKER_PROJ,
+        MODEL_TENSOR.A_GEN_FLOW_INPUT_PROJ,
+        MODEL_TENSOR.A_GEN_FLOW_COND_EMBD,
+        MODEL_TENSOR.A_GEN_FLOW_TIME_FREQS,
+        MODEL_TENSOR.A_GEN_FLOW_TIME_UP,
+        MODEL_TENSOR.A_GEN_FLOW_TIME_DOWN,
+        MODEL_TENSOR.A_GEN_FLOW_TIME_NORM,
+        MODEL_TENSOR.A_GEN_FLOW_BLK_NORM,
+        MODEL_TENSOR.A_GEN_FLOW_BLK_UP,
+        MODEL_TENSOR.A_GEN_FLOW_BLK_DOWN,
+        MODEL_TENSOR.A_GEN_FLOW_BLK_ADA,
+        MODEL_TENSOR.A_GEN_FLOW_FINAL_ADA,
+        MODEL_TENSOR.A_GEN_FLOW_FINAL_PROJ,
+        MODEL_TENSOR.A_GEN_OUT_EOS,
+        MODEL_TENSOR.A_GEN_INPUT_LINEAR,
+        MODEL_TENSOR.A_GEN_EMB_MEAN,
+        MODEL_TENSOR.A_GEN_EMB_STD,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_OUT,
+        MODEL_TENSOR.A_GEN_WAV_UPSAMPLE,
+        MODEL_TENSOR.A_GEN_WAV_SEANET_CONV_IN,
+        MODEL_TENSOR.A_GEN_WAV_SEANET_CONV_OUT,
+        MODEL_TENSOR.A_GEN_WAV_SEANET_RES_CONV1,
+        MODEL_TENSOR.A_GEN_WAV_SEANET_RES_CONV2,
+        MODEL_TENSOR.A_GEN_WAV_SEANET_SCALE_CONV,
         MODEL_TENSOR.A_ENC_CONV_NORM_MEAN,
         MODEL_TENSOR.A_ENC_CONV_NORM_VAR,
         MODEL_TENSOR.A_ENC_MEL_FILTERS,
@@ -4628,6 +4726,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.D2T,
     ],
     MODEL_ARCH.DFLASH: [
+        MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.ATTN_NORM,
         MODEL_TENSOR.ATTN_Q,
@@ -4903,6 +5002,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.POCKETTTS: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
 }
 
 # tensors that will not be serialized
@@ -5179,6 +5290,8 @@ class VisionProjectorType:
     NEMOTRON_V2_VL = "nemotron_v2_vl"
     QWEN3TTS_SPKENC = "qwen3tts_spkenc" # audio: ECAPA-TDNN speaker encoder
     QWEN3TTS_GEN = "qwen3tts_gen" # audio generation: code_predictor
+    POCKETTTS_SPKENC = "pockettts_spkenc" # audio: mimi encoder as voice-prompt encoder
+    POCKETTTS_GEN = "pockettts_gen" # audio generation: flow-matching decoder + mimi decoder
     HUNYUANVL      = "hunyuanvl"
     PARAKEET       = "parakeet"  # audio
     MINIMAXM3      = "minimax_m3"

gguf-py/gguf/gguf_writer.py

@@ -1453,6 +1453,9 @@ class GGUFWriter:
     def add_gen_audio_attention_layernorm_eps(self, value: float) -> None:
         self.add_float32(Keys.ClipGenAudio.Attention.LAYERNORM_EPS, value)
 
+    def add_gen_audio_model_variant(self, value: str) -> None:
+        self.add_string(Keys.ClipGenAudio.MODEL_VARIANT, value)
+
     def add_xielu_alpha_p(self, values: Sequence[float]):
         self.add_array(Keys.xIELU.ALPHA_P, values)
 

include/llama.h

@@ -203,11 +203,12 @@ extern "C" {
     };
 
     enum llama_load_mode {
-        LLAMA_LOAD_MODE_NONE       = 0, // no special loading mode
-        LLAMA_LOAD_MODE_MMAP       = 1, // memory map the model
-        LLAMA_LOAD_MODE_MLOCK      = 2, // force system to keep model in RAM rather than swapping or compressing
-        LLAMA_LOAD_MODE_MMAP_MLOCK = 3, // mmap + force system to keep model in RAM rather than swapping or compressing
-        LLAMA_LOAD_MODE_DIRECT_IO  = 4, // use direct I/O if available
+        LLAMA_LOAD_MODE_AUTO       = -1, // auto-detect based on device capabilities
+        LLAMA_LOAD_MODE_NONE       =  0, // no special loading mode
+        LLAMA_LOAD_MODE_MMAP       =  1, // memory map the model
+        LLAMA_LOAD_MODE_MLOCK      =  2, // force system to keep model in RAM rather than swapping or compressing
+        LLAMA_LOAD_MODE_MMAP_MLOCK =  3, // mmap + force system to keep model in RAM rather than swapping or compressing
+        LLAMA_LOAD_MODE_DIRECT_IO  =  4, // use direct I/O if available
     };
 
     LLAMA_API const char * llama_load_mode_name(enum llama_load_mode load_mode);
@@ -1324,8 +1325,6 @@ extern "C" {
     LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
     LLAMA_API void                   llama_sampler_reset (      struct llama_sampler * smpl);
     LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl);
-    // copy mutable sampler state without changing dst or its sampling graph bindings
-    // src and dst must have the same type and configuration
     LLAMA_API void                   llama_sampler_copy  (const struct llama_sampler * src, struct llama_sampler * dst);
     // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add)
     LLAMA_API void                   llama_sampler_free  (      struct llama_sampler * smpl);

models/templates/muse-glimmer.jinja

@@ -0,0 +1,211 @@
+{#
+  Template: Muse Glimmer ATEM Chat Template
+  Renders the ATEM tool-calling protocol: reasoning channel (to=self), tool
+  channels (to=<tool>), and the user channel, plus tool definitions and the
+  valid-recipient list in the system block.
+
+  Whitespace note: every tag uses the {%- -%} / {{- -}} stripping markers, so
+  the indentation below is purely for readability and contributes nothing to
+  the rendered output.
+#}
+{%- macro render_content(content) -%}
+    {%- if content is string -%}
+        {{- content -}}
+    {%- elif content is not none -%}
+        {%- for part in content -%}
+            {%- if part['type'] == 'image' -%}
+                {{- '<|patch|>' -}}
+            {%- elif part['type'] == 'video' -%}
+                {{- '<|video|>' -}}
+            {%- elif part['type'] == 'text' -%}
+                {{- part['text'] -}}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro render_atem(tc) -%}
+    {%- set args = tc.function.arguments -%}
+    {%- if args is not mapping -%}
+        {{- raise_exception('Muse Glimmer ATEM chat template requires tool_call.function.arguments to be a dict (mapping); a JSON string cannot be parsed in the HF jinja sandbox.') -}}
+    {%- endif -%}
+    {{- '<atem:function_calls>\n<atem:invoke name="' + tc.function.name + '">\n' -}}
+    {%- for k, v in args.items() -%}
+        {{- '<atem:parameter name="' + k + '">' -}}
+        {%- if v is boolean -%}
+            {%- if v -%}
+                true
+            {%- else -%}
+                false
+            {%- endif -%}
+        {%- elif v is none -%}
+            null
+        {%- elif v is mapping or (v is iterable and v is not string) -%}
+            {{- v | tojson -}}
+        {%- else -%}
+            {{- v -}}
+        {%- endif -%}
+        {{- '</atem:parameter>\n' -}}
+    {%- endfor -%}
+    {{- '</atem:invoke>\n</atem:function_calls>' -}}
+{%- endmacro -%}
+{%- macro render_tool_defs(tools) -%}
+    {{- 'In this environment you have access to a set of tools you can use to answer the user\'s question.\n\n' -}}
+    {{- 'You can invoke a function by writing a "<atem:function_calls>" block like the following:\n' -}}
+    {{- '<atem:function_calls>\n<atem:invoke name="$FUNCTION_NAME">\n<atem:parameter name="$PARAMETER_NAME">$PARAMETER_VALUE</atem:parameter>\n...\n</atem:invoke>\n</atem:function_calls>\n\n' -}}
+    {{- 'String and scalar parameters should be specified as is, while lists and objects should use JSON format. Note that spaces for string values are not stripped. The output is not expected to be valid XML and is parsed with regular expressions.\n' -}}
+    {{- 'Here are the functions available in JSONSchema format:\n' -}}
+    {{- '// Tool metadata\n' -}}
+    {%- set nsns = namespace(seen=[]) -%}
+    {%- for tool in tools -%}
+        {%- set fn = tool.function if tool.function is defined else tool -%}
+        {%- set tns = fn.name.split('.')[0] -%}
+        {%- if tns not in nsns.seen -%}
+            {%- set nsns.seen = nsns.seen + [tns] -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {%- set nd = tool_namespace_descriptions if tool_namespace_descriptions is defined else {} -%}
+    {%- for tns in nsns.seen -%}
+        {{- '{"name": ' + (tns | tojson) + ', "description": ' + ((nd[tns] if tns in nd else '') | tojson) + '}\n' -}}
+    {%- endfor -%}
+    {{- '// Function schemas' -}}
+    {%- for tool in tools -%}
+        {%- set fn = tool.function if tool.function is defined else tool -%}
+        {{- '\n{"name": ' + (fn.name | tojson) + ', "description": ' + (fn.description | tojson) + ', "parameters": ' + (fn.parameters | tojson) + '}' -}}
+    {%- endfor -%}
+    {{- '\n\nHere\'s an example of how to call a function in the tool set:\n' -}}
+    {{- '(If the tool namespace is not specified, invoke the function directly as `example_function_name` rather than `example_tool_name.example_function_name`)\n\n' -}}
+    {{- 'to=example_tool_name.example_function_name\n\n' -}}
+    {{- '<atem:function_calls>\n<atem:invoke name="example_tool_name.example_function_name">\n' -}}
+    {{- '<atem:parameter name="example_parameter_1">value_1</atem:parameter>\n' -}}
+    {{- '<atem:parameter name="example_parameter_2">This is the value for the second parameter\nthat can span\n"multiple" lines\n</atem:parameter>\n' -}}
+    {{- '</atem:invoke>\n</atem:function_calls>' -}}
+{%- endmacro -%}
+{%- macro render_reasoning() -%}
+    {%- set rs = reasoning_strength if reasoning_strength is defined and reasoning_strength else 'high' -%}
+    {{- 'Reasoning strength: ' + rs + '.' -}}
+{%- endmacro -%}
+{%- macro render_system_meta(tools) -%}
+    {%- set rns = namespace(recipients=['"self"'], nslist=[]) -%}
+    {%- if tools -%}
+        {%- for tool in tools -%}
+            {%- set fn = tool.function if tool.function is defined else tool -%}
+            {%- set tns = fn.name.split('.')[0] -%}
+            {%- if tns not in rns.nslist -%}
+                {%- set rns.nslist = rns.nslist + [tns] -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {%- for tns in rns.nslist -%}
+            {%- set rns.recipients = rns.recipients + ['"' + tns + '.*"'] -%}
+        {%- endfor -%}
+    {%- endif -%}
+    {%- set rns.recipients = rns.recipients + ['"user"'] -%}
+    {{- '# Valid recipients: ' + rns.recipients | join(', ') + '.' -}}
+{%- endmacro -%}
+{{- bos_token -}}
+{%- set ns = namespace(has_system=false) -%}
+{%- for m in messages -%}
+    {%- if m['role'] == 'system' -%}
+        {%- set ns.has_system = true -%}
+    {%- endif -%}
+{%- endfor -%}
+{%- if not ns.has_system -%}
+    {{- '<|start|>system<|message|>You are a helpful AI assistant.' -}}
+    {%- set kc = knowledge_cutoff if knowledge_cutoff is defined and knowledge_cutoff else '2026-01-04' -%}
+    {{- '\nKnowledge cutoff: ' + kc + '.' -}}
+    {%- if current_date is defined and current_date -%}
+        {{- '\nCurrent date: ' + current_date + '.' -}}
+    {%- elif strftime_now is defined -%}
+        {{- '\nCurrent date: ' + strftime_now('%Y-%m-%d') + '.' -}}
+    {%- endif -%}
+    {{- '\n\n' -}}
+    {{- render_reasoning() -}}
+    {%- if tools -%}
+        {{- '\n\n' -}}
+        {{- render_tool_defs(tools) -}}
+    {%- endif -%}
+    {{- '\n\n' -}}
+    {{- render_system_meta(tools) -}}
+    {{- '<|eot|>' -}}
+{%- endif -%}
+{%- for message in messages -%}
+    {%- set role = message['role'] -%}
+    {%- set end_token = '<|eom|>' if (not loop.last and messages[loop.index0 + 1]['role'] == role) else '<|eot|>' -%}
+    {%- if role == 'system' -%}
+        {#- Callers sometimes write the directive into the system prompt themselves.
+            Normalise "Reasoning effort" to "Reasoning strength" (jinja has no
+            case-insensitive replace, hence the four realistic casings), then skip
+            the kwarg-driven line below if the prompt already carries one. -#}
+        {%- set sys_text = render_content(message['content'])
+              | replace('Reasoning effort', 'Reasoning strength')
+              | replace('Reasoning Effort', 'Reasoning Strength')
+              | replace('reasoning effort', 'reasoning strength')
+              | replace('REASONING EFFORT', 'REASONING STRENGTH') -%}
+        {{- '<|start|>system<|message|>' -}}
+        {{- sys_text -}}
+        {%- if 'reasoning strength' not in (sys_text | lower) -%}
+            {{- '\n\n' -}}
+            {{- render_reasoning() -}}
+        {%- endif -%}
+        {%- if tools -%}
+            {{- '\n\n' -}}
+            {{- render_tool_defs(tools) -}}
+        {%- endif -%}
+        {{- '\n\n' -}}
+        {{- render_system_meta(tools) -}}
+        {{- '<|eot|>' -}}
+    {%- elif role == 'user' -%}
+        {{- '<|start|>user<|message|>' -}}
+        {{- render_content(message['content']) -}}
+        {{- '<|eot|>' -}}
+    {%- elif role == 'tool' -%}
+        {%- set tname = message.get('name') -%}
+        {%- if not tname -%}
+            {%- set tcid = message.get('tool_call_id') -%}
+            {%- set rns = namespace(name=tcid if tcid else '') -%}
+            {%- for m in messages -%}
+                {%- if m.get('tool_calls') -%}
+                    {%- for tc in m['tool_calls'] -%}
+                        {%- if tcid is not none and tc.id is defined and tc.id == tcid -%}
+                            {%- set rns.name = tc.function.name -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                {%- endif -%}
+            {%- endfor -%}
+            {%- set tname = rns.name -%}
+        {%- endif -%}
+        {{- '<|start|>tool ' + tname + '<|message|><tool_output name="' + tname + '">\n' -}}
+        {{- render_content(message['content']) -}}
+        {{- '\n</tool_output><|eot|>' -}}
+    {%- elif role == 'assistant' -%}
+        {%- if message.get('reasoning_content') -%}
+            {{- '<|start|>assistant to=self<|message|>' + message['reasoning_content'] + '<|eom|>' -}}
+        {%- endif -%}
+        {%- if message.get('tool_calls') -%}
+            {%- for tc in message['tool_calls'] -%}
+                {{- '<|start|>assistant to=' + tc.function.name + '<|message|>' -}}
+                {{- render_atem(tc) -}}
+                {%- if loop.last -%}
+                    {{- end_token -}}
+                {%- else -%}
+                    {{- '<|eom|>' -}}
+                {%- endif -%}
+            {%- endfor -%}
+        {%- else -%}
+            {%- set recipient = message.get('recipient') or 'user' -%}
+            {%- set end_turn = message.get('end_turn') -%}
+            {%- if end_turn is none -%}
+                {%- set end_turn = not (recipient and recipient != 'user') -%}
+            {%- endif -%}
+            {{- '<|start|>assistant' -}}
+            {%- if recipient -%}
+                {{- ' to=' + recipient -}}
+            {%- endif -%}
+            {{- '<|message|>' -}}
+            {{- render_content(message['content']) -}}
+            {{- ('<|eot|>' if end_turn else '<|eom|>') -}}
+        {%- endif -%}
+    {%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{- '<|start|>assistant' -}}
+{%- endif -%}

requirements/requirements-convert_hf_to_gguf.txt

@@ -2,8 +2,4 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 
 ## Embedding Gemma requires PyTorch 2.6.0 or later, bumped to 2.11.0 for compatibility
-torch==2.11.0; platform_machine != "s390x"
-
-# torch s390x packages can only be found from nightly builds
---extra-index-url https://download.pytorch.org/whl/nightly
-torch>=0.0.0.dev0; platform_machine == "s390x"
+torch==2.11.0

requirements/requirements-convert_lora_to_gguf.txt

@@ -1,4 +1,2 @@
 -r ./requirements-convert_hf_to_gguf.txt
 --extra-index-url https://download.pytorch.org/whl/cpu
-# torch s390x packages can only be found from nightly builds
---extra-index-url https://download.pytorch.org/whl/nightly

scripts/fetch_server_test_models.py

@@ -1,105 +0,0 @@
-#!/usr/bin/env python
-'''
-    This script fetches all the models used in the server tests.
-
-    This is useful for slow tests that use larger models, to avoid them timing out on the model downloads.
-
-    It is meant to be run from the root of the repository.
-
-    Example:
-        python scripts/fetch_server_test_models.py
-        ( cd tools/server/tests && ./tests.sh -v -x -m slow )
-'''
-import ast
-import glob
-import logging
-import os
-from typing import Generator
-from pydantic import BaseModel
-from typing import Optional
-import subprocess
-
-
-class HuggingFaceModel(BaseModel):
-    hf_repo: str
-    hf_file: Optional[str] = None
-
-    class Config:
-        frozen = True
-
-
-def collect_hf_model_test_parameters(test_file) -> Generator[HuggingFaceModel, None, None]:
-    try:
-        with open(test_file) as f:
-            tree = ast.parse(f.read())
-    except Exception as e:
-        logging.error(f'collect_hf_model_test_parameters failed on {test_file}: {e}')
-        return
-
-    for node in ast.walk(tree):
-        if isinstance(node, ast.FunctionDef):
-            for dec in node.decorator_list:
-                if isinstance(dec, ast.Call) and isinstance(dec.func, ast.Attribute) and dec.func.attr == 'parametrize':
-                    param_names = ast.literal_eval(dec.args[0]).split(",")
-                    if "hf_repo" not in param_names:
-                        continue
-
-                    raw_param_values = dec.args[1]
-                    if not isinstance(raw_param_values, ast.List):
-                        logging.warning(f'Skipping non-list parametrize entry at {test_file}:{node.lineno}')
-                        continue
-
-                    hf_repo_idx = param_names.index("hf_repo")
-                    hf_file_idx = param_names.index("hf_file") if "hf_file" in param_names else None
-
-                    for t in raw_param_values.elts:
-                        if not isinstance(t, ast.Tuple):
-                            logging.warning(f'Skipping non-tuple parametrize entry at {test_file}:{node.lineno}')
-                            continue
-                        yield HuggingFaceModel(
-                            hf_repo=ast.literal_eval(t.elts[hf_repo_idx]),
-                            hf_file=ast.literal_eval(t.elts[hf_file_idx]) if hf_file_idx is not None else None)
-
-
-if __name__ == '__main__':
-    logging.basicConfig(level=logging.INFO, format='%(levelname)s: %(message)s')
-
-    models = sorted(list(set([
-        model
-        for test_file in glob.glob('tools/server/tests/unit/test_*.py')
-        for model in collect_hf_model_test_parameters(test_file)
-    ])), key=lambda m: (m.hf_repo, m.hf_file))
-
-    logging.info(f'Found {len(models)} models in parameterized tests:')
-    for m in models:
-        logging.info(f'  - {m.hf_repo} / {m.hf_file}')
-
-    cli_path = os.environ.get(
-        'LLAMA_CLI_BIN_PATH',
-        os.path.join(
-            os.path.dirname(__file__),
-            '../build/bin/Release/llama-cli.exe' if os.name == 'nt' else '../build/bin/llama-cli'))
-
-    for m in models:
-        if '<' in m.hf_repo or (m.hf_file is not None and '<' in m.hf_file):
-            continue
-        if m.hf_file is not None and '-of-' in m.hf_file:
-            logging.warning(f'Skipping model at {m.hf_repo} / {m.hf_file} because it is a split file')
-            continue
-        logging.info(f'Using llama-cli to ensure model {m.hf_repo}/{m.hf_file} was fetched')
-        cmd = [
-            cli_path,
-            '-hfr', m.hf_repo,
-            *([] if m.hf_file is None else ['-hff', m.hf_file]),
-            '-n', '1',
-            '-p', 'Hey',
-            '--no-warmup',
-            '--log-disable',
-            '-st']
-        if m.hf_file != 'tinyllamas/stories260K.gguf' and 'Mistral-Nemo' not in m.hf_repo:
-            cmd += ('-fa', 'on')
-        try:
-            subprocess.check_call(cmd)
-        except subprocess.CalledProcessError:
-            logging.error(f'Failed to fetch model at {m.hf_repo} / {m.hf_file} with command:\n  {" ".join(cmd)}')
-            exit(1)

scripts/hip/gcn-cdna-vgpr-check.py

@@ -60,90 +60,10 @@ def main():
     log_file = sys.argv[1]
     ignored = {
         '_ZL21gated_linear_attn_f32ILi128EEviiiifPKfS1_S1_S1_S1_Pf',
-        '_ZL18flash_attn_ext_f16ILi64ELi64ELi16ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi16ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi16ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi64ELi64ELi32ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
         '_ZL13rwkv_wkv7_f32ILi128EEviiiiPKfS1_S1_S1_S1_S1_S1_Pf',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi16ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi16ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi32ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi16ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi2ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi32ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi16ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi32ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi1ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi2ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi2ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi2ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi2ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi2ELi8ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi16ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi16ELi4ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi32ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi4ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi4ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi4ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi4ELi4ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi4ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi4ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi64ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi64ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi64ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi64ELi1ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi64ELi64ELi8ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi8ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi8ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi8ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi8ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi4ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi8ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi8ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi2ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi8ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi8ELi8ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL24mul_mat_q_stream_k_fixupIL9ggml_type22ELi8ELb1EEvPKiS2_PfPKfiiimimimi',
-        '_ZL9mul_mat_qIL9ggml_type3ELi32ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type3ELi48ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type20ELi32ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type17ELi64ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL18flash_attn_ext_f16ILi80ELi80ELi4ELi4ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL15flash_attn_tileILi256ELi256ELi32ELi1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL9mul_mat_qIL9ggml_type19ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type17ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type22ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type19ELi128ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type19ELi128ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type7ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type3ELi128ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type3ELi128ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type7ELi128ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type7ELi128ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type11ELi112ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type11ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL24mul_mat_q_stream_k_fixupIL9ggml_type11ELi128ELb0EEvPKiS2_PfPKfiiimimimi',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi32ELi1ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL9mul_mat_qIL9ggml_type2ELi112ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi32ELi2ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi112ELi112ELi4ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi32ELi1ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi32ELi2ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi128ELi128ELi4ELi8ELb1ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_f16ILi96ELi96ELi4ELi8ELb0ELb0EEvPKcS1_S1_S1_S1_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS5_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL18flash_attn_ext_vecILi128ELi2EL9ggml_type2ELS0_2ELb0EEvPKcS2_S2_S2_S2_PKiPfP15HIP_vector_typeIfLj2EEffffjfiS6_IjLj3EEiiiiiiiiiiiliiliiiiil',
-        '_ZL9mul_mat_qIL9ggml_type10ELi16ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type12ELi128ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type40ELi112ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type40ELi112ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type40ELi128ELb0EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii',
-        '_ZL9mul_mat_qIL9ggml_type40ELi128ELb1EEvPKcPKiS4_S4_PfS5_iiiiiiiiiiiiiiiii'
+        '_ZL12rwkv_wkv_f32ILi128EEviiiiPKfS1_S1_S1_S1_S1_Pf',
+        '_ZL9mul_mat_qIL9ggml_type10ELi64ELb1EEvPKcPKiS4_S4_PfS5_PKf15HIP_vector_typeIjLj3EEiiiiiS9_S9_iiiS9_S9_iiiS9_',
+        '_ZL9mul_mat_qIL9ggml_type42ELi128ELb1EEvPKcPKiS4_S4_PfS5_PKf15HIP_vector_typeIjLj3EEiiiiiS9_S9_iiiS9_S9_iiiS9_',
     }
 
     functions = parse_log_file(log_file)

src/llama-arch.cpp

@@ -147,6 +147,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MELLUM,           "mellum"           },
     { LLM_ARCH_NANBEIGE,         "nanbeige"         },
     { LLM_ARCH_QWEN3TTS,         "qwen3tts"         },
+    { LLM_ARCH_POCKETTTS,        "pockettts"        },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 

src/llama-arch.h

@@ -152,6 +152,7 @@ enum llm_arch {
     LLM_ARCH_DFLASH,
     LLM_ARCH_NANBEIGE,
     LLM_ARCH_QWEN3TTS,
+    LLM_ARCH_POCKETTTS,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-model-loader.cpp

@@ -543,7 +543,7 @@ llama_model_loader::llama_model_loader(
 
     tensor_buft_overrides = param_tensor_buft_overrides_p;
 
-    this->use_mmap      = load_mode == LLAMA_LOAD_MODE_MMAP || load_mode == LLAMA_LOAD_MODE_MMAP_MLOCK;
+    this->use_mmap      = load_mode == LLAMA_LOAD_MODE_MMAP || load_mode == LLAMA_LOAD_MODE_MMAP_MLOCK || load_mode == LLAMA_LOAD_MODE_AUTO;
     this->use_direct_io = load_mode == LLAMA_LOAD_MODE_DIRECT_IO;
 
     if (!fname.empty()) {

src/llama-model.cpp

@@ -116,6 +116,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_qwen3vlmoe(params);
         case LLM_ARCH_QWEN3TTS:
             return new llama_model_qwen3tts(params);
+        case LLM_ARCH_POCKETTTS:
+            return new llama_model_pockettts(params);
         case LLM_ARCH_PHI2:
             return new llama_model_phi2(params);
         case LLM_ARCH_PHI3:
@@ -1271,8 +1273,23 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
 
     this->ml = &ml; // to be used by create_tensor() and load_arch_tensors()
 
+    if (ml.use_mmap && params.load_mode == LLAMA_LOAD_MODE_AUTO) {
+        for (const auto & dev : devices) {
+            ggml_backend_dev_props props;
+            ggml_backend_dev_get_props(dev.dev, &props);
+            if (!props.caps.mmap_support) {
+                ml.use_mmap = false;
+                break;
+            }
+        }
+    }
+
+    const char * load_mode_name = params.load_mode == LLAMA_LOAD_MODE_AUTO
+        ? llama_load_mode_name(ml.use_mmap ? LLAMA_LOAD_MODE_MMAP : LLAMA_LOAD_MODE_NONE)
+        : llama_load_mode_name(params.load_mode);
+
     LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (load_mode = %s)\n",
-        __func__, llama_load_mode_name(params.load_mode));
+        __func__, load_mode_name);
 
     // build a list of buffer types for the CPU and GPU devices
     pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
@@ -2452,7 +2469,7 @@ llama_model_params llama_model_default_params() {
         /*.tensor_buft_overrides       =*/ nullptr,
         /*.n_gpu_layers                =*/ -1,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
-        /*.load_mode                   =*/ LLAMA_LOAD_MODE_MMAP,
+        /*.load_mode                   =*/ LLAMA_LOAD_MODE_AUTO,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
@@ -2622,6 +2639,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MAINCODER:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_NANBEIGE:
+        case LLM_ARCH_POCKETTTS:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

src/llama-model.h

@@ -623,8 +623,9 @@ struct llama_model {
     struct ggml_tensor * per_layer_model_proj = nullptr;
     struct ggml_tensor * per_layer_proj_norm  = nullptr;
 
-    // eagle3
-    struct ggml_tensor * fc  = nullptr;  // feature fusion layer
+    // eagle3 / dflash feature fusion layer
+    struct ggml_tensor * fc   = nullptr;
+    struct ggml_tensor * fc_s = nullptr;
     struct ggml_tensor * d2t = nullptr;  // draft to target vocabulary mapping
 
     // dspark

src/llama.cpp

@@ -48,6 +48,8 @@ const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_ty
 
 const char * llama_load_mode_name(enum llama_load_mode load_mode) {
     switch (load_mode) {
+        case LLAMA_LOAD_MODE_AUTO:
+            return "auto";
         case LLAMA_LOAD_MODE_NONE:
             return "none";
         case LLAMA_LOAD_MODE_MMAP:
@@ -63,11 +65,12 @@ const char * llama_load_mode_name(enum llama_load_mode load_mode) {
 }
 
 enum llama_load_mode llama_load_mode_from_str(const char * str) {
-    if (std::strcmp(str, "none") == 0)       { return LLAMA_LOAD_MODE_NONE;       }
-    if (std::strcmp(str, "mmap") == 0)       { return LLAMA_LOAD_MODE_MMAP;       }
-    if (std::strcmp(str, "mlock") == 0)      { return LLAMA_LOAD_MODE_MLOCK;      }
+    if (std::strcmp(str, "auto")       == 0) { return LLAMA_LOAD_MODE_AUTO;       }
+    if (std::strcmp(str, "none")       == 0) { return LLAMA_LOAD_MODE_NONE;       }
+    if (std::strcmp(str, "mmap")       == 0) { return LLAMA_LOAD_MODE_MMAP;       }
+    if (std::strcmp(str, "mlock")      == 0) { return LLAMA_LOAD_MODE_MLOCK;      }
     if (std::strcmp(str, "mmap+mlock") == 0) { return LLAMA_LOAD_MODE_MMAP_MLOCK; }
-    if (std::strcmp(str, "dio") == 0)        { return LLAMA_LOAD_MODE_DIRECT_IO;  }
+    if (std::strcmp(str, "dio")        == 0) { return LLAMA_LOAD_MODE_DIRECT_IO;  }
     throw std::invalid_argument(std::string("unknown load mode: ") + str);
 }
 

src/models/dflash.cpp

@@ -79,6 +79,7 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
 
     const int64_t n_embd_inp = hparams.n_embd_inp_enc();
 
+    tok_embd        = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,       "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
     // DSpark = DFlash + a semi-autoregressive Markov head and Confidence head
     //
     // TODO: only Qwen3-style backbones are supported for now; other backbones (e.g. Gemma4)
@@ -97,6 +98,7 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
     }
 
     fc              = create_tensor(tn(LLM_TENSOR_FC,              "weight"), { n_embd_inp, n_embd }, 0);
+    fc_s            = create_tensor(tn(LLM_TENSOR_FC,              "scale"),  { 1 }, TENSOR_NOT_REQUIRED);
     output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
     output_norm     = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM,    "weight"), { n_embd }, 0); // decoder final norm
 
@@ -205,7 +207,7 @@ template <>
 llama_model_dflash::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     ggml_tensor * cur = build_inp_embd_enc();
 
-    cur = build_lora_mm(model.fc, cur);
+    cur = build_lora_mm(model.fc, cur, model.fc_s);
     cb(cur, "fc_out", -1);
 
     cur = build_norm(cur, model.output_norm_enc, NULL, LLM_NORM_RMS, -1);
@@ -460,9 +462,9 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
         cb(cur, "ffn_norm", il);
 
         cur = build_ffn(cur,
-                layer.ffn_up,   NULL, NULL,
-                layer.ffn_gate, NULL, NULL,
-                layer.ffn_down, NULL, NULL,
+                layer.ffn_up,   NULL, layer.ffn_up_s,
+                layer.ffn_gate, NULL, layer.ffn_gate_s,
+                layer.ffn_down, NULL, layer.ffn_down_s,
                 NULL,
                 LLM_FFN_SILU, LLM_FFN_PAR, il);
         cb(cur, "ffn_out", il);
@@ -479,15 +481,17 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
     res->t_embd = cur;
 
     // lm_head from the target model (shared via ctx_other)
-    auto * output = model.output;
+    auto * output   = model.output;
+    auto * output_s = model.output_s;
     if (output == nullptr) {
         GGML_ASSERT(cparams.ctx_other != nullptr);
         const auto * model_other = llama_get_model(cparams.ctx_other);
         GGML_ASSERT(model_other->output != nullptr && "DFlash decoder requires the target model's output projection");
-        output = model_other->output;
+        output   = model_other->output;
+        output_s = model_other->output_s;
     }
 
-    cur = build_lora_mm(output, cur);
+    cur = build_lora_mm(output, cur, output_s);
     cb(cur, "result_output", -1);
     res->t_logits = cur;
 
@@ -655,15 +659,17 @@ llama_model_dflash::graph_dsv4::graph_dsv4(const llama_model & model, const llm_
     cb(cur, "result_norm", -1);
 
     // lm_head from the target model (shared via ctx_other)
-    auto * output = model.output;
+    auto * output   = model.output;
+    auto * output_s = model.output_s;
     if (output == nullptr) {
         GGML_ASSERT(cparams.ctx_other != nullptr);
         const auto * model_other = llama_get_model(cparams.ctx_other);
         GGML_ASSERT(model_other->output != nullptr && "DSpark decoder requires the target model's output projection");
-        output = model_other->output;
+        output   = model_other->output;
+        output_s = model_other->output_s;
     }
 
-    cur = build_lora_mm(output, cur);
+    cur = build_lora_mm(output, cur, output_s);
     cb(cur, "result_output", -1);
     res->t_logits = cur;
 

src/models/exaone4.cpp

@@ -1,6 +1,9 @@
 #include "models.h"
 
 void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");
+
     if (hparams.n_layer() == 64) {    // 32B
         hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
         hparams.n_swa = 4096;
@@ -15,9 +18,6 @@ void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.n_layer_nextn, false);
-
-    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");
 
     switch (hparams.n_layer()) {
         case 30: type = LLM_TYPE_1_2B; break;

src/models/models.h

@@ -713,6 +713,19 @@ struct llama_model_gpt2 : public llama_model_base {
 };
 
 
+struct llama_model_pockettts : public llama_model_base {
+    llama_model_pockettts(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_codeshell : public llama_model_base {
     llama_model_codeshell(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;

src/models/nemotron-h.cpp

@@ -177,8 +177,11 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
     auto * inp = build_inp_mem_hybrid();
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
+    const bool extract_final_inp = (size_t) n_layer < cparams.embeddings_layer_inp.size() && cparams.embeddings_layer_inp[n_layer];
 
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         struct ggml_tensor * inpSA = inpL;
 
         // norm
@@ -195,7 +198,7 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
             cur = build_ffn_layer(cur, model, il);
         }
 
-        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
+        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked && !extract_final_inp) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -209,6 +212,13 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
     }
 
     cur = inpL;
+    if (extract_final_inp) {
+        res->t_layer_inp[n_layer] = cur;
+
+        if (inp_out_ids && cparams.embeddings_nextn_masked) {
+            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+        }
+    }
 
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 

src/models/pockettts.cpp

@@ -0,0 +1,146 @@
+#include "models.h"
+
+// backbone of the pocket-tts CALM pipeline: the "text" side of a flow language model.
+// it has no lm_head, the audio latents are produced by the flow net inside the mmproj
+
+void llama_model_pockettts::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+    switch (hparams.n_layer()) {
+        case 6:  type = LLM_TYPE_109M; break;
+        case 24: type = LLM_TYPE_335M; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_pockettts::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    // no output head, the logits are unused; reuse the embedding table so a sampler can still run
+    output        = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, TENSOR_NOT_REQUIRED);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_pockettts::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_pockettts::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur, model.output_s);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

tests/peg-parser/test-json-parser.cpp

@@ -77,6 +77,30 @@ void test_json_parser(testing &t) {
         t.assert_equal("result_is_need_more_input", true, result.need_more_input());
     });
 
+    // Test need_more_input() parsing - incomplete escape sequence in a string value
+    t.test("need_more_input() parsing - incomplete escape sequence", [](testing &t) {
+        auto json = build_peg_parser([](common_peg_parser_builder & p) { return p.json(); });
+
+        std::vector<std::string> inputs {
+            R"({"text": "hello\)",     // dangling backslash
+            R"({"text": "hello\u)",    // incomplete unicode escape sequence
+            R"({"text": "hello\u00)",
+        };
+
+        for (const auto & input : inputs) {
+            t.test(input, [&](testing &t) {
+                common_peg_parse_context ctx(input, COMMON_PEG_PARSE_FLAG_LENIENT);
+
+                auto result = json.parse(ctx);
+
+                t.assert_equal("result_is_need_more_input", true, result.need_more_input());
+
+                // the incomplete escape sequence is not part of the partial value
+                t.assert_equal("result_end", input.find('\\'), result.end);
+            });
+        }
+    });
+
     t.test("object member", [](testing &t) {
         auto parser = build_peg_parser([](common_peg_parser_builder & p) {
             return p.json_member("name", "\"" + p.chars("[a-z]") + "\"");

tests/test-backend-ops.cpp

@@ -8827,6 +8827,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 128, 4));
 
     test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 1));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 4));
     test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 1));
     test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 4));
     test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 128, 4));
@@ -9011,6 +9012,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 128, k, {12,1}, {1,1}));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 50, 200, k));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, true, 50, 200, k));
+        test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_BF16, GGML_TYPE_F32, 16, 16, false, 50, 200, k));
+        test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_BF16, GGML_TYPE_F32, 16, 16, true, 50, 200, k));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F32, GGML_TYPE_F32, 16, 16, false, 50, 200, k));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F32, GGML_TYPE_F32, 16, 16, true, 50, 200, k));
     }
@@ -9042,6 +9045,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 2, 2, b, 32, 8192, 64));
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 50, 200, 64));
+        test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_BF16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
+        test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_BF16, GGML_TYPE_F32, 16, 16, b, 50, 200, 64));
     }
 
     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));

tests/test-backend-sampler.cpp

@@ -2105,7 +2105,9 @@ static std::vector<const backend_test_case *> collect_tests_to_run(const std::st
 #ifdef GGML_USE_HIP
             // TODO: remove this when https://github.com/ggml-org/llama.cpp/pull/26592 is merged
             if (test.name == "penalties" || test.name == "set_sampler" ||
-                test.name == "mixed"     || test.name == "top_p") {
+                test.name == "mixed"     || test.name == "top_p"       ||
+                test.name == "multi_output_sampling_chain" ||
+                test.name == "multi_output_cpu") {
                 fprintf(stderr, "Skipping test '%s' on HIP backend (no backend TOP_K support)\n", test.name.c_str());
                 continue;
             }

tests/test-chat.cpp

@@ -5843,6 +5843,52 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
+    // Muse Glimmer format tests
+    {
+        auto tst = peg_tester("models/templates/muse-glimmer.jinja", detailed_debug);
+
+        const std::string call_markup =
+            "<atem:function_calls>\n"
+            "<atem:invoke name=\"special_function\">\n"
+            "<atem:parameter name=\"arg1\">1</atem:parameter>\n"
+            "</atem:invoke>\n"
+            "</atem:function_calls>";
+
+        // A plain answer is unaffected
+        tst.test(" to=user<|message|>Hello, world!\nWhat's up?<|eot|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(message_assist)
+            .run();
+
+        // "Inform then act": the model answers the user and calls a tool in ONE generation,
+        // closing the answer with <|eom|>. The answer must stop there rather than swallow it.
+        tst.test(" to=user<|message|>Hello, world!\nWhat's up?<|eom|>"
+                 "<|start|>assistant to=special_function<|message|>" +
+                 call_markup)
+            .tools({ special_function_tool })
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(message_with_content_and_tool_call("Hello, world!\nWhat's up?", "special_function",
+                                                       "{\"arg1\":1}"))
+            .run();
+
+        // Markup quoted in an answer has no preceding <|eom|>, so it stays content instead of
+        // becoming an invocation the user never asked for
+        tst.test(" to=user<|message|>You invoke it like this:\n" + call_markup + "<|eot|>")
+            .tools({ special_function_tool })
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect_content("You invoke it like this:\n" + call_markup)
+            .run();
+
+        // Tool markup inside the analysis channel is reasoning, not a call
+        tst.test(" to=self<|message|>I could use " + call_markup + " here<|eom|>"
+                 "<|start|>assistant to=user<|message|>Hello!<|eot|>")
+            .tools({ special_function_tool })
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect_reasoning("I could use " + call_markup + " here")
+            .expect_content("Hello!")
+            .run();
+    }
+
     // GPT-OSS format tests
     {
         auto tst = peg_tester("models/templates/openai-gpt-oss-120b.jinja", detailed_debug);

tools/cli/README.md

@@ -58,7 +58,7 @@
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |

tools/completion/README.md

@@ -141,7 +141,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |

tools/imatrix/imatrix.cpp

@@ -222,6 +222,15 @@ static void compute_cossim(std::vector<tensor_statistics> & tstats) {
     }
 }
 
+static bool all_finite(const float * v, size_t n) {
+    for (size_t i = 0; i < n; ++i) {
+        if (!std::isfinite(v[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
     GGML_UNUSED(user_data);
 
@@ -299,33 +308,39 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             exit(1); //GGML_ABORT("fatal error");
         }
         LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
-        // loop over all possible experts, regardless if they are used or not in the batch
-        for (int64_t ex = 0; ex < n_as; ++ex) {
-            size_t e_start = ex*src1->ne[0];
 
-            for (int64_t idx = 0; idx < n_ids; ++idx) {
-                for (int64_t row = 0; row < src1->ne[2]; ++row) {
-                    const int excur = *(const int32_t *) (m_ids.data() + row*ids->nb[1] + idx*ids->nb[0]);
+        const int64_t ne0      = src1->ne[0];
+        const int64_t n_tokens = src1->ne[2];
 
-                    GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
+        // single pass over the routing ids
+        std::vector<uint8_t> touched(n_as, 0);
+        for (int64_t idx = 0; idx < n_ids; ++idx) {
+            for (int64_t row = 0; row < n_tokens; ++row) {
+                const int32_t ex = *(const int32_t *) (m_ids.data() + row * ids->nb[1] + idx * ids->nb[0]);
 
-                    if (excur != ex) continue;
+                GGML_ASSERT(ex >= 0 && ex < n_as);  // sanity check
 
-                    const int64_t i11 = idx % src1->ne[1];
-                    const int64_t i12 = row;
-                    const float * x = (const float *)(data + i11*src1->nb[1] + i12*src1->nb[2]);
+                const int64_t i11 = idx % src1->ne[1];
+                const float * x   = (const float *) (data + i11 * src1->nb[1] + row * src1->nb[2]);
+                float *       acc = e.values.data() + ex * ne0;
 
-                    e.counts[ex]++;
-
-                    for (int64_t j = 0; j < src1->ne[0]; ++j) {
-                        e.values[e_start + j] += x[j] * x[j];
-                        if (!std::isfinite((float)e.values[e_start + j])) {
-                            LOG_ERR("%f detected in %s\n", (float)e.values[e_start + j], wname.c_str());
-                            exit(1);
-                        }
-                    }
+                e.counts[ex]++;
+                touched[ex] = 1;
+                for (int64_t j = 0; j < ne0; ++j) {
+                    acc[j] += x[j] * x[j];
                 }
             }
+        }
+
+        // check for non-finite values, only checking experts that were routed to and touched
+        for (int64_t ex = 0; ex < n_as; ++ex) {
+            if (touched[ex] && !all_finite(e.values.data() + ex * ne0, ne0)) {
+                LOG_ERR("%s: non-finite values detected in %s\n", __func__, wname.c_str());
+                exit(1);
+            }
+        }
+
+        for (int64_t ex = 0; ex < n_as; ++ex) {
             const int32_t n_chunk = e.counts[ex] / chunk_size;
             if (n_chunk > m_last_chunk) {
                 const int32_t chunk_step = n_chunk - m_last_chunk;
@@ -366,24 +381,28 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->ne[2], (int)src1->type);
 
+        const int64_t ne0 = src1->ne[0];
+
         for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {
             for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {
                 // handle 3D+ tensors, but flatten 3D+ activations when model tensor is 2D
                 const int64_t mat_id = (i3 % src0->ne[3]) * src0->ne[2] + (i2 % src0->ne[2]);
-                const int64_t mat_start = mat_id * src1->ne[0];
+                float *       acc    = e.values.data() + mat_id * ne0;
 
                 for (int64_t row = 0; row < src1->ne[1]; ++row) {
                     const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]);
-                    for (int64_t j = 0; j < src1->ne[0]; ++j) {
-                        e.values[mat_start + j] += x[j] * x[j];
-                        if (!std::isfinite((float)e.values[j])) {
-                            LOG_ERR("%f detected in %s\n", (float)e.values[j], wname.c_str());
-                            exit(1);
-                        }
+                    for (int64_t j = 0; j < ne0; ++j) {
+                        acc[j] += x[j] * x[j];
                     }
                 }
             }
         }
+
+        // check for non-finite values
+        if (!all_finite(e.values.data(), e.values.size())) {
+            LOG_ERR("%s: non-finite values detected in %s\n", __func__, wname.c_str());
+            exit(1);
+        }
         // only 1 count in practice, except when a tensor is used for both MUL_MAT_ID and MUL_MAT
         for (size_t i = 0; i < e.counts.size(); ++i) {
             e.counts[i] += ggml_nrows(src1) / n_mat;

tools/llama-bench/llama-bench.cpp

@@ -384,7 +384,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_gpu_layers         */ { -1 },
     /* n_cpu_moe            */ { 0 },
     /* split_mode           */ { LLAMA_SPLIT_MODE_LAYER },
-    /* load_mode            */ { LLAMA_LOAD_MODE_MMAP },
+    /* load_mode            */ { LLAMA_LOAD_MODE_AUTO },
     /* main_gpu             */ { 0 },
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { LLAMA_FLASH_ATTN_TYPE_AUTO },
@@ -459,7 +459,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -nkvo, --no-kv-offload <0|1>                      (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -fa, --flash-attn <on|off|auto>                   (default: %s)\n", join(transform_to_str(cmd_params_defaults.flash_attn, llama_flash_attn_type_name), ",").c_str());
     printf("  -dev, --device <dev0/dev1/...>                    (default: auto)\n");
-    printf("  -lm, --load-mode <none|mmap|mlock|mmap+mlock|dio> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
+    printf("  -lm, --load-mode <auto|none|mmap|mlock|mmap+mlock|dio> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
     printf("  -mmp, --mmap <0|1>                                (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -dio, --direct-io <0|1>                           (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -embd, --embeddings <0|1>                         (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
@@ -764,7 +764,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 std::vector<llama_load_mode> modes;
                 for (const auto & m : p) {
                     llama_load_mode mode;
-                    if (m == "none") {
+                    if (m == "auto") {
+                        mode = LLAMA_LOAD_MODE_AUTO;
+                    } else if (m == "none") {
                         mode = LLAMA_LOAD_MODE_NONE;
                     } else if (m == "mmap") {
                         mode = LLAMA_LOAD_MODE_MMAP;

tools/mtmd/CMakeLists.txt

@@ -57,6 +57,9 @@ add_library(mtmd
             models/mimo-audio.cpp
             models/qwen3tts-spkenc.cpp
             models/qwen3tts-gen.cpp
+            models/pockettts-seanet.cpp
+            models/pockettts-spkenc.cpp
+            models/pockettts-gen.cpp
             models/step3vl.cpp
             models/siglip.cpp
             models/whisper-enc.cpp

tools/mtmd/README-dev.md

@@ -59,8 +59,10 @@ Due to wide variety of audio generation pipelines, the `mtmd_gen_audio` system i
 
 ### Checklist for porting new audio generation models to mtmd
 
-1. Establish a list of reusable and missing components from the current mtmd implementation.
-2. For GGUF conversion:
+1. Make sure to consult merged PRs about adding new TTS models, especially reviewer comments
+    - Example: https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+mtmd+tts+is%3Amerged
+2. Establish a list of reusable and missing components from the current mtmd implementation.
+3. For GGUF conversion:
     - Backbone model should be converted to a normal text model (loadable via `libllama`)
         - If model used hard-coded embedding row ID, append them to token embeddings and assign token name for them (see `qwen3tts.py`)
         - If model have a specific output logits head for audio codes (usually semantic code), keep the head as-is and pad the logits at inference time (see `src/models/qwen3vl.cpp`)
@@ -70,12 +72,17 @@ Due to wide variety of audio generation pipelines, the `mtmd_gen_audio` system i
     - For tensor naming:
         - Prefixed with `a.*` for tensors used by speaker encoder pipeline
         - Prefixed with `a.gen.*` for generation stages (code / mel-spectrogram / PCM generation)
-3. Make sure most of the changes happen inside `mtmd-helper-gen.cpp`. A good PR looks like this:
+    - For GGUF metadata:
+        - Reuse as many existing keys as possible
+        - In most cases, you can hard-code model configs in the model graph class, or in `clip_hparams`
+        - If some values need to be exposed to the `mtmd_helper` layer, hard-code them in `mtmd_helper` and distinguish by pipeline and `mtmd_gen_audio_info::model_variant` if necessary
+        - Do NOT add new GGUF metadata or new fields to `mtmd_gen_audio_info` unless you can prove that you absolutely need them
+4. Make sure most of the changes happen inside `mtmd-helper-gen.cpp`. A good PR looks like this:
     - 10-20% changes is to add new backbone (text) model and conversion
     - 60% changes inside `mtmd-helper-gen.cpp`
     - 10% changes inside `libmtmd` and `clip.cpp` systems
     - The rest downstream code (CLI, server) should have no changes at all
-4. Update usage documentation in `tools/tts/README.md`
+5. Update usage documentation in `tools/tts/README.md`
 
 IMPORTANT: If your model needs changes that don't fit the existing infrastructure, **open an issue first for discussion**.
 

tools/mtmd/clip-impl.h

@@ -92,7 +92,9 @@
 #define KEY_A_LOCAL_GROUP_SIZE     "clip.audio.local_group_size"     // mimo-v2.5: input_local_transformer grouping size
 // audio generation (gen-audio)-specific
 #define KEY_GEN_AUDIO_PROJ_TYPE    "clip.gen.audio.projector_type" // for models with mixed modalities
-#define KEY_AUDIO_SUBSAMPLING_FACTOR "clip.audio.subsampling_factor"
+// name of the weight variant, for settings that are not in the checkpoint
+#define KEY_GEN_AUDIO_VARIANT      "clip.gen.audio.model_variant"
+#define KEY_AUDIO_SUBSMPL_FACTOR   "clip.audio.subsampling_factor"
 
 //
 // tensor name constants
@@ -246,6 +248,38 @@
 #define TN_A_GEN_WAV_DAC_POST_SNAKE  "a.gen.wav.dac.post_snake.%s"
 #define TN_A_GEN_WAV_DAC_POST_CONV   "a.gen.wav.dac.post_conv.%s"
 
+// pocket-tts
+#define TN_A_SEANET_CONV_IN      "a.seanet.conv_in.%s"
+#define TN_A_SEANET_CONV_OUT     "a.seanet.conv_out.%s"
+#define TN_A_SEANET_RES_CONV1    "a.seanet.blk.%d.res_conv1.%s"
+#define TN_A_SEANET_RES_CONV2    "a.seanet.blk.%d.res_conv2.%s"
+#define TN_A_SEANET_SCALE_CONV   "a.seanet.blk.%d.scale_conv.%s"
+#define TN_A_SPEAKER_PROJ        "a.speaker_proj.%s"
+#define TN_A_DOWNSAMPLE_CONV     "a.downsample.conv.%s"
+#define TN_A_GEN_FLOW_INPUT_PROJ "a.gen.flow.input_proj.%s"
+#define TN_A_GEN_FLOW_COND_EMBD  "a.gen.flow.cond_embd.%s"
+#define TN_A_GEN_FLOW_TIME_FREQS "a.gen.flow.time.%d.freqs"
+#define TN_A_GEN_FLOW_TIME_UP    "a.gen.flow.time.%d.up.%s"
+#define TN_A_GEN_FLOW_TIME_DOWN  "a.gen.flow.time.%d.down.%s"
+#define TN_A_GEN_FLOW_TIME_NORM  "a.gen.flow.time.%d.norm"
+#define TN_A_GEN_FLOW_BLK_NORM   "a.gen.flow.blk.%d.norm.%s"
+#define TN_A_GEN_FLOW_BLK_UP     "a.gen.flow.blk.%d.up.%s"
+#define TN_A_GEN_FLOW_BLK_DOWN   "a.gen.flow.blk.%d.down.%s"
+#define TN_A_GEN_FLOW_BLK_ADA    "a.gen.flow.blk.%d.ada.%s"
+#define TN_A_GEN_FLOW_FINAL_ADA  "a.gen.flow.final.ada.%s"
+#define TN_A_GEN_FLOW_FINAL_PROJ "a.gen.flow.final.proj.%s"
+#define TN_A_GEN_OUT_EOS         "a.gen.out_eos.%s"
+#define TN_A_GEN_INPUT_LINEAR    "a.gen.input_linear.%s"
+#define TN_A_GEN_EMB_MEAN        "a.gen.emb_mean"
+#define TN_A_GEN_EMB_STD         "a.gen.emb_std"
+#define TN_A_GEN_WAV_QUANT_OUT   "a.gen.wav.quant_out.%s"
+#define TN_A_GEN_WAV_UPSAMPLE    "a.gen.wav.upsample.%s"
+#define TN_A_GEN_WAV_SEANET_CONV_IN    "a.gen.wav.seanet.conv_in.%s"
+#define TN_A_GEN_WAV_SEANET_CONV_OUT   "a.gen.wav.seanet.conv_out.%s"
+#define TN_A_GEN_WAV_SEANET_RES_CONV1  "a.gen.wav.seanet.blk.%d.res_conv1.%s"
+#define TN_A_GEN_WAV_SEANET_RES_CONV2  "a.gen.wav.seanet.blk.%d.res_conv2.%s"
+#define TN_A_GEN_WAV_SEANET_SCALE_CONV "a.gen.wav.seanet.blk.%d.scale_conv.%s"
+
 // cogvlm
 #define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
 #define TN_MM_H_TO_4H      "mm.up.%s"
@@ -455,6 +489,8 @@ enum projector_type {
     PROJECTOR_TYPE_MIMO_AUDIO,
     PROJECTOR_TYPE_QWEN3TTS_SPKENC,
     PROJECTOR_TYPE_QWEN3TTS_GEN,
+    PROJECTOR_TYPE_POCKETTTS_SPKENC,
+    PROJECTOR_TYPE_POCKETTTS_GEN,
     PROJECTOR_TYPE_MUSE_GLIMMER,
     PROJECTOR_TYPE_UNKNOWN,
 };
@@ -515,6 +551,8 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_PARAKEET,          "parakeet"},
     { PROJECTOR_TYPE_QWEN3TTS_SPKENC,   "qwen3tts_spkenc"},
     { PROJECTOR_TYPE_QWEN3TTS_GEN,      "qwen3tts_gen"},
+    { PROJECTOR_TYPE_POCKETTTS_SPKENC,  "pockettts_spkenc"},
+    { PROJECTOR_TYPE_POCKETTTS_GEN,     "pockettts_gen"},
     { PROJECTOR_TYPE_MUSE_GLIMMER,      "muse-glimmer"},
 };
 

tools/mtmd/clip-model.h

@@ -141,6 +141,20 @@ struct clip_hparams {
     int32_t rvq_num_quantizers = 0;
     std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)
 
+    // threshold for the "out_eos_score" graph output
+    float gen_eos_threshold = 0.0f;
+
+    // name of the weight variant, some pipelines tune themselves on it
+    std::string gen_model_variant;
+
+    // pocket-tts
+    static constexpr int32_t pockettts_max_spk_seconds = 30;
+    int32_t seanet_n_stage    = 0;
+    std::vector<int32_t> seanet_ratios; // encoder order (reversed compared to the config)
+    int32_t mimi_downsample   = 0;      // encoder frame rate / model frame rate
+    int32_t mimi_tfm_context  = 0;      // attention window of the mimi transformers, in frames
+    int32_t flow_n_step       = 1;      // lsd_decode steps
+
     // qwen3tts code2wav
     int32_t wav_tfm_n_layer      = 0;
     int32_t wav_tfm_n_embd       = 0;
@@ -402,6 +416,63 @@ struct qf_block {
     std::vector<clip_layer> qf_proj_layers;
 };
 
+// pocket-tts SEANet stack, used in both directions:
+// encoder = conv_in -> per stage (residual unit, strided conv) -> conv_out
+// decoder = conv_in -> per stage (strided convtr, residual unit) -> conv_out
+struct clip_seanet {
+    // one residual unit: ELU -> dilated conv -> ELU -> pointwise conv, added to the input
+    struct stage {
+        ggml_tensor * res_conv1_w = nullptr;
+        ggml_tensor * res_conv1_b = nullptr;
+        ggml_tensor * res_conv2_w = nullptr;
+        ggml_tensor * res_conv2_b = nullptr;
+        ggml_tensor * scale_conv_w = nullptr; // strided conv (encoder) or convtr (decoder)
+        ggml_tensor * scale_conv_b = nullptr;
+    };
+
+    ggml_tensor * conv_in_w  = nullptr;
+    ggml_tensor * conv_in_b  = nullptr;
+    ggml_tensor * conv_out_w = nullptr;
+    ggml_tensor * conv_out_b = nullptr;
+    std::vector<stage> stages;
+};
+
+// pocket-tts flow-matching decoder (SimpleMLPAdaLN)
+struct clip_flow_net {
+    // AdaLN res block: in_ln -> modulate -> Linear -> SiLU -> Linear, gated residual
+    struct block {
+        ggml_tensor * norm_w = nullptr;
+        ggml_tensor * norm_b = nullptr;
+        ggml_tensor * up_w   = nullptr;
+        ggml_tensor * up_b   = nullptr;
+        ggml_tensor * down_w = nullptr;
+        ggml_tensor * down_b = nullptr;
+        ggml_tensor * ada_w  = nullptr; // -> shift, scale, gate
+        ggml_tensor * ada_b  = nullptr;
+    };
+
+    // timestep embedder: cos/sin(t * freqs) -> Linear -> SiLU -> Linear -> RMSNorm
+    struct time_embd {
+        ggml_tensor * freqs  = nullptr;
+        ggml_tensor * up_w   = nullptr;
+        ggml_tensor * up_b   = nullptr;
+        ggml_tensor * down_w = nullptr;
+        ggml_tensor * down_b = nullptr;
+        ggml_tensor * norm   = nullptr; // RMSNorm alpha
+    };
+
+    ggml_tensor * input_proj_w = nullptr;
+    ggml_tensor * input_proj_b = nullptr;
+    ggml_tensor * cond_embd_w  = nullptr;
+    ggml_tensor * cond_embd_b  = nullptr;
+    ggml_tensor * final_ada_w  = nullptr; // -> shift, scale
+    ggml_tensor * final_ada_b  = nullptr;
+    ggml_tensor * final_proj_w = nullptr;
+    ggml_tensor * final_proj_b = nullptr;
+    std::vector<time_embd> time;
+    std::vector<block> blocks;
+};
+
 // qwen3tts code2wav: RVQ codes -> raw PCM
 struct clip_code2wav {
     // "upsample" stage: one ConvNeXt block plus the causal ConvTranspose1d before it
@@ -699,6 +770,24 @@ struct clip_model {
     // qwen3tts code2wav: RVQ codes -> raw PCM
     clip_code2wav c2w;
 
+    // pocket-tts: SEANet stack, shared by the encoder (speaker path) and the decoder (gen path)
+    clip_seanet seanet;
+
+    // pocket-tts: voice latent -> backbone embd (speaker path)
+    ggml_tensor * spk_proj_w      = nullptr;
+    ggml_tensor * downsample_w    = nullptr;
+
+    // pocket-tts: flow-matching decoder, backbone hidden state -> next latent
+    clip_flow_net flow;
+    ggml_tensor * gen_out_eos_w   = nullptr;
+    ggml_tensor * gen_out_eos_b   = nullptr;
+    ggml_tensor * gen_input_lin_w = nullptr; // latent -> backbone embd
+    ggml_tensor * gen_emb_mean    = nullptr;
+    ggml_tensor * gen_emb_std     = nullptr;
+    ggml_tensor * gen_quant_out_w = nullptr; // latent -> decoder dim
+    ggml_tensor * gen_upsample_w  = nullptr; // depthwise convtr, frame rate -> encoder frame rate
+    std::vector<clip_layer> gen_tfm_layers; // mimi decoder_transformer
+
     // cogvlm
     ggml_tensor * mm_post_fc_norm_w = nullptr;
     ggml_tensor * mm_post_fc_norm_b = nullptr;

tools/mtmd/clip.cpp

@@ -174,6 +174,10 @@ struct clip_ctx {
 
     bool support_batch = false;
 
+    // for audio gen, reseeded only when the caller asks for another seed
+    std::mt19937 rng{std::random_device{}()};
+    uint32_t rng_seed = UINT32_MAX;
+
     clip_ctx(clip_context_params & ctx_params) {
         flash_attn_type = ctx_params.flash_attn_type;
         no_alloc = ctx_params.no_alloc;
@@ -1059,6 +1063,25 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
             {
                 builder = std::make_unique<clip_graph_qwen3tts_spkenc>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+            {
+                builder = std::make_unique<clip_graph_pockettts_spkenc>(ctx, img);
+            } break;
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            {
+                const auto gen_process = params ? params->gen_process : CLIP_GEN_PROCESS_GEN_CODE;
+                const int  n_step = ctx->model.hparams.flow_n_step;
+                const int64_t n_latent = ctx->model.gen_input_lin_w->ne[0];
+                GGML_ASSERT(n_step > 0);
+                GGML_ASSERT(n_latent > 0);
+                // "inp_feats" takes the caller's buffer as-is, the graph must consume all of it
+                if (params && params->feats) {
+                    GGML_ASSERT(params->feats->size() % (size_t) n_latent == 0);
+                    GGML_ASSERT(params->feats->size() >= (size_t) n_latent);
+                }
+                const int  n_frames = params && params->feats ? (int) (params->feats->size() / n_latent) : 1;
+                builder = std::make_unique<clip_graph_pockettts_gen>(ctx, img, gen_process, n_step, n_frames);
+            } break;
         case PROJECTOR_TYPE_QWEN3TTS_GEN:
             {
                 const auto  gen_process = params ? params->gen_process : CLIP_GEN_PROCESS_GEN_CODE;
@@ -1282,6 +1305,7 @@ struct clip_model_loader {
                 // these are unused, but still need to be set to avoid issues
                 hparams.image_size = 0;
                 hparams.patch_size = 1;
+                get_string(KEY_GEN_AUDIO_VARIANT, hparams.gen_model_variant, false);
 
             } else {
                 GGML_ASSERT(false && "unknown modality");
@@ -1421,7 +1445,7 @@ struct clip_model_loader {
                     } break;
                 case PROJECTOR_TYPE_PARAKEET:
                     {
-                        get_u32(KEY_AUDIO_SUBSAMPLING_FACTOR, hparams.subsampling_factor);
+                        get_u32(KEY_AUDIO_SUBSMPL_FACTOR, hparams.subsampling_factor);
                         GGML_ASSERT(hparams.subsampling_factor == 8 &&
                             "subsampling_factor must match the conv strides in clip_graph_parakeet::build()");
                         get_u32(KEY_A_CONV_KERNEL_SIZE,       hparams.audio_conv_kernel_size);
@@ -1745,6 +1769,22 @@ struct clip_model_loader {
                         // matches the reference decoder's sliding_window (speech_tokenizer/config.json)
                         hparams.wav_tfm_swa = 72;
                     } break;
+                case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+                case PROJECTOR_TYPE_POCKETTTS_GEN:
+                    {
+                        // mimi front-end takes the raw waveform, no mel
+                        hparams.audio_sample_rate = 24000;
+                        // seanet ratios are [6,5,4] in the config, the encoder reverses them
+                        hparams.seanet_ratios   = { 4, 5, 6 };
+                        hparams.seanet_n_stage  = (int32_t) hparams.seanet_ratios.size();
+                        hparams.mimi_downsample = 16;
+                        // matches the reference transformer's "context"
+                        hparams.mimi_tfm_context = 250;
+                        hparams.rope_theta       = 10000.0f;
+                        // flow_lm defaults, see pocket_tts/default_parameters.py
+                        hparams.flow_n_step       = 1;
+                        hparams.gen_eos_threshold = -4.0f;
+                    } break;
                 case PROJECTOR_TYPE_PADDLEOCR:
                     {
                         hparams.n_merge = 2;
@@ -1947,7 +1987,9 @@ struct clip_model_loader {
 
                 // GEMMA4UA is encoder-free: it uses n_mel_bins as a raw-waveform frame size (640) and has no FFT/filterbank, so the mel-range and FFT
                 // checks below do not apply to it.
-                const bool fft_based = model.proj_type != PROJECTOR_TYPE_GEMMA4UA;
+                // pocket-tts is encoder-free in the same sense: mimi convolves the raw waveform
+                const bool fft_based = model.proj_type != PROJECTOR_TYPE_GEMMA4UA &&
+                                       model.proj_type != PROJECTOR_TYPE_POCKETTTS_SPKENC;
 
                 // Validate audio hparams loaded from GGUF metadata
                 if (hparams.n_mel_bins <= 0 || (fft_based && hparams.n_mel_bins > 256)) {
@@ -2020,6 +2062,31 @@ struct clip_model_loader {
             return cur;
         };
 
+        // pocket-tts: the encoder and the decoder share the same layout, only the prefix differs
+        auto load_seanet = [&](clip_seanet & seanet, bool is_decoder) {
+            const char * conv_in  = is_decoder ? TN_A_GEN_WAV_SEANET_CONV_IN    : TN_A_SEANET_CONV_IN;
+            const char * conv_out = is_decoder ? TN_A_GEN_WAV_SEANET_CONV_OUT   : TN_A_SEANET_CONV_OUT;
+            const char * res1     = is_decoder ? TN_A_GEN_WAV_SEANET_RES_CONV1  : TN_A_SEANET_RES_CONV1;
+            const char * res2     = is_decoder ? TN_A_GEN_WAV_SEANET_RES_CONV2  : TN_A_SEANET_RES_CONV2;
+            const char * scale    = is_decoder ? TN_A_GEN_WAV_SEANET_SCALE_CONV : TN_A_SEANET_SCALE_CONV;
+
+            seanet.conv_in_w  = get_tensor(string_format(conv_in,  "weight"));
+            seanet.conv_in_b  = get_tensor(string_format(conv_in,  "bias"));
+            seanet.conv_out_w = get_tensor(string_format(conv_out, "weight"));
+            seanet.conv_out_b = get_tensor(string_format(conv_out, "bias"));
+
+            seanet.stages.resize(hparams.seanet_n_stage);
+            for (int i = 0; i < hparams.seanet_n_stage; i++) {
+                auto & stage = seanet.stages[i];
+                stage.res_conv1_w  = get_tensor(string_format(res1,  i, "weight"));
+                stage.res_conv1_b  = get_tensor(string_format(res1,  i, "bias"));
+                stage.res_conv2_w  = get_tensor(string_format(res2,  i, "weight"));
+                stage.res_conv2_b  = get_tensor(string_format(res2,  i, "bias"));
+                stage.scale_conv_w = get_tensor(string_format(scale, i, "weight"));
+                stage.scale_conv_b = get_tensor(string_format(scale, i, "bias"));
+            }
+        };
+
         auto get_vector = [&](const std::string & name) {
             std::vector<float> result;
             auto it = tensor_offset.find(name);
@@ -2081,7 +2148,8 @@ struct clip_model_loader {
 
         const bool has_standard_layers = (
             model.proj_type != PROJECTOR_TYPE_GEMMA3NV &&
-            model.proj_type != PROJECTOR_TYPE_QWEN3TTS_SPKENC);
+            model.proj_type != PROJECTOR_TYPE_QWEN3TTS_SPKENC &&
+            model.proj_type != PROJECTOR_TYPE_POCKETTTS_GEN);
 
         // layers
         const int n_layers_to_load = has_standard_layers ? hparams.n_layer : 0;
@@ -2755,6 +2823,81 @@ struct clip_model_loader {
                     model.mm_fc_w = get_tensor(string_format(TN_MM_AUDIO_FC, "weight"));
                     model.mm_fc_b = get_tensor(string_format(TN_MM_AUDIO_FC, "bias"));
                 } break;
+            case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+                {
+                    load_seanet(model.seanet, false);
+                    model.downsample_w = get_tensor(string_format(TN_A_DOWNSAMPLE_CONV, "weight"));
+                    model.spk_proj_w   = get_tensor(string_format(TN_A_SPEAKER_PROJ, "weight"));
+                } break;
+            case PROJECTOR_TYPE_POCKETTTS_GEN:
+                {
+                    auto & flow = model.flow;
+                    flow.input_proj_w = get_tensor(string_format(TN_A_GEN_FLOW_INPUT_PROJ, "weight"));
+                    flow.input_proj_b = get_tensor(string_format(TN_A_GEN_FLOW_INPUT_PROJ, "bias"));
+                    flow.cond_embd_w  = get_tensor(string_format(TN_A_GEN_FLOW_COND_EMBD, "weight"));
+                    flow.cond_embd_b  = get_tensor(string_format(TN_A_GEN_FLOW_COND_EMBD, "bias"));
+                    flow.final_ada_w  = get_tensor(string_format(TN_A_GEN_FLOW_FINAL_ADA, "weight"));
+                    flow.final_ada_b  = get_tensor(string_format(TN_A_GEN_FLOW_FINAL_ADA, "bias"));
+                    flow.final_proj_w = get_tensor(string_format(TN_A_GEN_FLOW_FINAL_PROJ, "weight"));
+                    flow.final_proj_b = get_tensor(string_format(TN_A_GEN_FLOW_FINAL_PROJ, "bias"));
+
+                    flow.time.resize(2);
+                    for (size_t i = 0; i < flow.time.size(); i++) {
+                        auto & t = flow.time[i];
+                        t.freqs  = get_tensor(string_format(TN_A_GEN_FLOW_TIME_FREQS, (int) i));
+                        t.up_w   = get_tensor(string_format(TN_A_GEN_FLOW_TIME_UP,   (int) i, "weight"));
+                        t.up_b   = get_tensor(string_format(TN_A_GEN_FLOW_TIME_UP,   (int) i, "bias"));
+                        t.down_w = get_tensor(string_format(TN_A_GEN_FLOW_TIME_DOWN, (int) i, "weight"));
+                        t.down_b = get_tensor(string_format(TN_A_GEN_FLOW_TIME_DOWN, (int) i, "bias"));
+                        t.norm   = get_tensor(string_format(TN_A_GEN_FLOW_TIME_NORM, (int) i));
+                    }
+
+                    // one AdaLN block per flow depth, the count is only known from the tensors
+                    for (int il = 0; ; il++) {
+                        ggml_tensor * probe = get_tensor(string_format(TN_A_GEN_FLOW_BLK_NORM, il, "weight"), false);
+                        if (probe == nullptr) {
+                            break;
+                        }
+                        clip_flow_net::block blk;
+                        blk.norm_w = probe;
+                        blk.norm_b = get_tensor(string_format(TN_A_GEN_FLOW_BLK_NORM, il, "bias"));
+                        blk.up_w   = get_tensor(string_format(TN_A_GEN_FLOW_BLK_UP,   il, "weight"));
+                        blk.up_b   = get_tensor(string_format(TN_A_GEN_FLOW_BLK_UP,   il, "bias"));
+                        blk.down_w = get_tensor(string_format(TN_A_GEN_FLOW_BLK_DOWN, il, "weight"));
+                        blk.down_b = get_tensor(string_format(TN_A_GEN_FLOW_BLK_DOWN, il, "bias"));
+                        blk.ada_w  = get_tensor(string_format(TN_A_GEN_FLOW_BLK_ADA,  il, "weight"));
+                        blk.ada_b  = get_tensor(string_format(TN_A_GEN_FLOW_BLK_ADA,  il, "bias"));
+                        flow.blocks.push_back(blk);
+                    }
+
+                    model.gen_out_eos_w   = get_tensor(string_format(TN_A_GEN_OUT_EOS, "weight"));
+                    model.gen_out_eos_b   = get_tensor(string_format(TN_A_GEN_OUT_EOS, "bias"));
+                    model.gen_input_lin_w = get_tensor(string_format(TN_A_GEN_INPUT_LINEAR, "weight"));
+                    model.gen_emb_mean    = get_tensor(TN_A_GEN_EMB_MEAN);
+                    model.gen_emb_std     = get_tensor(TN_A_GEN_EMB_STD);
+
+                    // mimi decoder
+                    model.gen_quant_out_w = get_tensor(string_format(TN_A_GEN_WAV_QUANT_OUT, "weight"));
+                    model.gen_upsample_w  = get_tensor(string_format(TN_A_GEN_WAV_UPSAMPLE, "weight"));
+                    load_seanet(model.seanet, true);
+                    model.gen_tfm_layers.resize(hparams.n_layer);
+                    for (int il = 0; il < hparams.n_layer; il++) {
+                        auto & layer = model.gen_tfm_layers[il];
+                        const char * p = "a.gen.wav.tfm";
+                        layer.ln_1_w    = get_tensor(string_format(TN_LN_1,        p, il, "weight"));
+                        layer.ln_1_b    = get_tensor(string_format(TN_LN_1,        p, il, "bias"));
+                        layer.q_w       = get_tensor(string_format(TN_ATTN_Q,      p, il, "weight"));
+                        layer.k_w       = get_tensor(string_format(TN_ATTN_K,      p, il, "weight"));
+                        layer.v_w       = get_tensor(string_format(TN_ATTN_V,      p, il, "weight"));
+                        layer.o_w       = get_tensor(string_format(TN_ATTN_OUTPUT, p, il, "weight"));
+                        layer.ls_1_w    = get_tensor(string_format(TN_LS_1,        p, il, "weight"));
+                        layer.ln_2_w    = get_tensor(string_format(TN_LN_2,        p, il, "weight"));
+                        layer.ln_2_b    = get_tensor(string_format(TN_LN_2,        p, il, "bias"));
+                        layer.ff_up_w   = get_tensor(string_format(TN_FFN_UP,      p, il, "weight"));
+                        layer.ff_down_w = get_tensor(string_format(TN_FFN_DOWN,    p, il, "weight"));
+                        layer.ls_2_w    = get_tensor(string_format(TN_LS_2,        p, il, "weight"));
+                    }
+                } break;
             case PROJECTOR_TYPE_QWEN3TTS_GEN:
                 {
                     // code_predictor
@@ -4060,6 +4203,17 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
                 // one hidden-state vector fed back to the talker per call
                 n_patches = 1;
             } break;
+        case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+            {
+                // one conditioning row per 12.5Hz frame
+                const int hop = ctx->model.hparams.mimi_downsample * 120;
+                n_patches = img->nx() / hop;
+            } break;
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            {
+                // one latent per call for GEN_CODE, GEN_WAV sizes its input from the caller
+                n_patches = 1;
+            } break;
         case PROJECTOR_TYPE_GRANITE4_VISION:
             {
                 // Per-tile output token count: each projector block outputs
@@ -4101,6 +4255,15 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
     return clip_encode(ctx, &params);
 }
 
+// persisted state slots of the gen-audio decoder, per pipeline
+static std::vector<c2w_state_slot> list_gen_state_slots(const clip_hparams & hparams, const clip_model & model) {
+    switch (model.proj_type) {
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:  return list_c2w_state_slots(hparams, model);
+        case PROJECTOR_TYPE_POCKETTTS_GEN: return list_pockettts_state_slots(hparams, model);
+        default:                           return {};
+    }
+}
+
 bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
     const clip_image_f32_batch & imgs = *params->imgs;
     int n_batch_cur = imgs.entries.size();
@@ -4116,6 +4279,11 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
         clip_model_loader::warmup(*ctx, *params->imgs);
     }
 
+    if (params->seed != ctx->rng_seed) {
+        ctx->rng_seed = params->seed;
+        ctx->rng.seed(params->seed == UINT32_MAX ? std::random_device{}() : params->seed);
+    }
+
     // build the inference graph
     ggml_backend_sched_reset(ctx->sched.get());
     ggml_cgraph * gf = clip_get_graph_builder(ctx, imgs, params)->build();
@@ -4160,6 +4328,50 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
         ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur));
     };
 
+    // upload the decoder state from the previous call, or zero-fill on a cold start
+    auto set_gen_state_in = [&]() {
+        size_t offset = 0;
+        for (const auto & slot : list_gen_state_slots(hparams, model)) {
+            ggml_tensor * t = get_inp_tensor(("state_in_" + slot.name).c_str());
+            const size_t nb = ggml_nbytes(t);
+            if (params->state_in && params->state_in->size() >= offset + nb) {
+                ggml_backend_tensor_set(t, params->state_in->data() + offset, 0, nb);
+            } else {
+                std::vector<uint8_t> zeros(nb, 0);
+                ggml_backend_tensor_set(t, zeros.data(), 0, nb);
+            }
+            offset += nb;
+        }
+    };
+
+    // rope positions and attention mask of the mimi transformers (pocket-tts).
+    // the mask is causal with a sliding window, see _build_attention_mask() in the reference
+    auto set_pockettts_tfm_inputs = [&]() {
+        const int64_t n_pos = ggml_nelements(get_inp_tensor("inp_pos"));
+        GGML_ASSERT(n_pos > 0);
+        std::vector<int32_t> positions((size_t) n_pos);
+        for (int64_t i = 0; i < n_pos; i++) {
+            positions[(size_t) i] = (int32_t) i;
+        }
+        set_input_i32("inp_pos", positions);
+
+        // the preprocessor truncates the waveform to keep this mask bounded
+        const int64_t max_pos = (int64_t) clip_hparams::pockettts_max_spk_seconds * hparams.audio_sample_rate / 120;
+        GGML_ASSERT(n_pos <= max_pos && "pocket-tts speaker reference too long for a dense mask");
+
+        const int64_t context = hparams.mimi_tfm_context;
+        std::vector<float> mask((size_t) n_pos * n_pos, -INFINITY);
+        for (int64_t q = 0; q < n_pos; q++) {
+            for (int64_t k = 0; k < n_pos; k++) {
+                const int64_t delta = q - k;
+                if (delta >= 0 && delta < context) {
+                    mask[(size_t) q * n_pos + k] = 0.0f;
+                }
+            }
+        }
+        set_input_f32("kq_mask", mask);
+    };
+
     // set input pixel values
     if (!imgs.is_audio) {
         size_t nelem = 0;
@@ -4203,8 +4415,8 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
         }
         set_input_f32("inp_raw", inp_raw);
 
-    } else if (!(ctx->proj_type() == PROJECTOR_TYPE_QWEN3TTS_GEN && params->gen_process == CLIP_GEN_PROCESS_GEN_WAV)) {
-        // audio input, code2wav is not here: its only input is "inp_codes", set in the switch below
+    } else if (params->gen_process != CLIP_GEN_PROCESS_GEN_WAV) {
+        // audio input. GEN_WAV is not here: it takes codes or feats, set in the switch below
         GGML_ASSERT(imgs.entries.size() == 1);
 
         const auto & mel_inp = imgs.entries[0];
@@ -4737,6 +4949,30 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
                 }
                 set_input_i32("patches", patches);
             } break;
+        case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+            {
+                set_pockettts_tfm_inputs();
+            } break;
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            {
+                if (params->gen_process == CLIP_GEN_PROCESS_GEN_WAV) {
+                    GGML_ASSERT(params->feats != nullptr);
+                    set_input_f32("inp_feats", *params->feats);
+                    // positions and mask are derived in-graph from the persisted counter
+                    set_gen_state_in();
+                } else {
+                    // flow matching starts from gaussian noise, std = sqrt(temp)
+                    ggml_tensor * t = get_inp_tensor("inp_noise");
+                    // Config.default_temperature, for a caller that does not set one
+                    const float temp = params->temp > 0.0f ? params->temp : 0.7f;
+                    std::normal_distribution<float> dist(0.0f, std::sqrt(temp));
+                    std::vector<float> noise(ggml_nelements(t));
+                    for (auto & v : noise) {
+                        v = dist(ctx->rng);
+                    }
+                    set_input_f32("inp_noise", noise);
+                }
+            } break;
         case PROJECTOR_TYPE_GEMMA4V:
         case PROJECTOR_TYPE_GEMMA4UV:
             {
@@ -4861,20 +5097,7 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
                         }
                     }
                     set_input_i32("inp_codes", codes);
-
-                    // upload the state from the previous call, or zero-fill on a cold start
-                    size_t offset = 0;
-                    for (const auto & slot : list_c2w_state_slots(hparams, model)) {
-                        ggml_tensor * t = get_inp_tensor(("state_in_" + slot.name).c_str());
-                        const size_t nb = ggml_nbytes(t);
-                        if (params->state_in && params->state_in->size() >= offset + nb) {
-                            ggml_backend_tensor_set(t, params->state_in->data() + offset, 0, nb);
-                        } else {
-                            std::vector<uint8_t> zeros(nb, 0);
-                            ggml_backend_tensor_set(t, zeros.data(), 0, nb);
-                        }
-                        offset += nb;
-                    }
+                    set_gen_state_in();
                 } else {
                     // code0 indexes gen_code_out_embd_w via ggml_get_rows; bound it
                     const int64_t vocab0 = model.gen_code_out_embd_w->ne[1];
@@ -4886,11 +5109,10 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
                     set_input_i32("inp_code0", code0);
 
                     // one uniform(0,1) draw per codebook, used by do_sampling()
-                    static std::mt19937 rng{ std::random_device{}() };
                     std::uniform_real_distribution<float> dist(0.0f, 1.0f);
                     const int64_t n_acoustic = model.gen_code_head_w->ne[2];
                     for (int64_t g = 0; g < n_acoustic; g++) {
-                        std::vector<float> r = { dist(rng) };
+                        std::vector<float> r = { dist(ctx->rng) };
                         set_input_f32(("inp_rand_" + std::to_string(g)).c_str(), r);
                     }
                 }
@@ -5343,14 +5565,31 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
     // for audio gen models
     //
 
+    // optional outputs: a pipeline yields codes or feats, and not all have an eos head
     if (params->out_codes != nullptr) {
         ggml_tensor * codes = ggml_graph_get_tensor(gf, "out_codes");
-        if (codes == nullptr) {
-            GGML_ABORT("out_codes requested but graph has no \"out_codes\" tensor");
+        if (codes != nullptr) {
+            auto & out_codes = *params->out_codes;
+            out_codes.resize(ggml_nelements(codes));
+            ggml_backend_tensor_get(codes, out_codes.data(), 0, ggml_nbytes(codes));
+        }
+    }
+    if (params->out_feats != nullptr) {
+        ggml_tensor * feats = ggml_graph_get_tensor(gf, "out_feats");
+        if (feats != nullptr) {
+            auto & out_feats = *params->out_feats;
+            out_feats.resize(ggml_nelements(feats));
+            ggml_backend_tensor_get(feats, out_feats.data(), 0, ggml_nbytes(feats));
+        }
+    }
+    if (params->out_is_eos != nullptr) {
+        ggml_tensor * eos = ggml_graph_get_tensor(gf, "out_eos_score");
+        if (eos != nullptr) {
+            GGML_ASSERT(ggml_nelements(eos) == 1);
+            float score = 0.0f;
+            ggml_backend_tensor_get(eos, &score, 0, sizeof(float));
+            *params->out_is_eos = score > hparams.gen_eos_threshold;
         }
-        auto & out_codes = *params->out_codes;
-        out_codes.resize(ggml_nelements(codes));
-        ggml_backend_tensor_get(codes, out_codes.data(), 0, ggml_nbytes(codes));
     }
     if (params->out_audio != nullptr) {
         ggml_tensor * audio = ggml_graph_get_tensor(gf, "out_audio");
@@ -5362,9 +5601,9 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
         ggml_backend_tensor_get(audio, out_audio.data(), 0, ggml_nbytes(audio));
 
         // drop the tail audio that comes from the code-0 rear padding
-        const int64_t n_codes    = model.gen_code_head_w->ne[2] + 1;
+        const int64_t n_codes    = params->codes ? model.gen_code_head_w->ne[2] + 1 : 0;
         const int64_t n_frames_w = hparams.wav_tfm_swa;
-        const int64_t n_frames   = (int64_t) params->codes->size() / n_codes;
+        const int64_t n_frames   = params->codes ? (int64_t) params->codes->size() / n_codes : n_frames_w;
         if (n_frames < n_frames_w) {
             const size_t hop = out_audio.size() / n_frames_w;
             out_audio.resize((size_t) n_frames * hop);
@@ -5373,12 +5612,12 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
     if (params->state_out != nullptr) {
         auto & state_out = *params->state_out;
         size_t total = 0;
-        for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+        for (const auto & slot : list_gen_state_slots(hparams, model)) {
             total += (size_t) (slot.ne0 * slot.ne1) * sizeof(float);
         }
         state_out.resize(total);
         size_t offset = 0;
-        for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+        for (const auto & slot : list_gen_state_slots(hparams, model)) {
             ggml_tensor * t = ggml_graph_get_tensor(gf, ("state_out_" + slot.name).c_str());
             if (t == nullptr) {
                 GGML_ABORT("state_out requested but graph has no \"state_out_%s\" tensor", slot.name.c_str());
@@ -5526,6 +5765,10 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_fc_w->ne[2];
         case PROJECTOR_TYPE_QWEN3TTS_GEN:
             return ctx->model.gen_code_out_embd_w->ne[0];
+        case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+            return ctx->model.spk_proj_w->ne[1];
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            return ctx->model.gen_input_lin_w->ne[1];
         case PROJECTOR_TYPE_PARAKEET:
             return ctx->model.mm_1_w->ne[1];
         default:

tools/mtmd/clip.h

@@ -104,9 +104,14 @@ struct clip_encode_params {
     int32_t top_k = 50;
     float   top_p = 1.0f;
     std::vector<int32_t> * out_codes = nullptr; // this frame's 16 sampled codes
+    std::vector<float> * out_feats = nullptr; // continuous counterpart of out_codes
+    uint32_t seed = UINT32_MAX;               // UINT32_MAX for random
+    float   temp = 0.0f;                      // sampling temperature, noise scale for flow-matching decoders
+    bool * out_is_eos = nullptr;
 
     // GEN_WAV
     const std::vector<int32_t> * codes = nullptr;     // this frame's 16 RVQ codes
+    const std::vector<float> *   feats = nullptr;     // continuous counterpart of codes
     std::vector<float> * out_audio = nullptr;         // decoded PCM samples, F32
     const std::vector<uint8_t> * state_in  = nullptr; // state from previous call, null or wrong size means cold start
     std::vector<uint8_t> *       state_out = nullptr; // state for the next call

tools/mtmd/models/models.h

@@ -318,6 +318,59 @@ struct clip_graph_qwen3tts_gen : clip_graph {
     };
 };
 
+//
+// pocket-tts: SEANet convolution stack, shared by the voice encoder and the mimi decoder.
+// stateless unless state_in is populated: convs then pad instead of carrying left-context.
+//
+struct clip_graph_pockettts_seanet : clip_graph {
+    clip_graph_pockettts_seanet(const clip_graph & parent) : clip_graph(parent) {}
+    ggml_cgraph * build() override { GGML_ABORT("call encode()/decode() instead"); }
+
+    // per-call streaming state, keyed by slot name (see list_pockettts_state_slots)
+    std::map<std::string, ggml_tensor *> state_in;
+    mutable std::vector<std::pair<std::string, ggml_tensor *>> state_out;
+
+    ggml_tensor * conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, int dilation,
+                         bool pad_replicate = false, const std::string & state_name = "") const;
+    ggml_tensor * conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride,
+                                   const std::string & state_name = "") const;
+    ggml_tensor * res_unit(ggml_tensor * x, const clip_seanet::stage & stage, int dilation,
+                           const std::string & state_prefix = "") const;
+
+    // x: [T, C] -> [T / hop, dim]
+    ggml_tensor * encode(ggml_tensor * x) const;
+    // x: [T, dim] -> [T * hop, 1], streams when state_in is populated
+    ggml_tensor * decode(ggml_tensor * x) const;
+};
+
+// mimi encoder + speaker_proj: reference waveform -> voice conditioning rows
+struct clip_graph_pockettts_spkenc : clip_graph {
+    clip_graph_pockettts_spkenc(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * tfm_layer_forward(ggml_tensor * cur, const clip_layer & layer, ggml_tensor * inp_pos, ggml_tensor * kq_mask, int il) const;
+};
+
+//
+// pocket-tts generation:
+// GEN_CODE = flow-matching decoder + end-of-speech head, one latent per call
+// GEN_WAV  = mimi decoder, a window of latents -> PCM
+//
+struct clip_graph_pockettts_gen : clip_graph {
+    clip_graph_pockettts_gen(clip_ctx * ctx, const clip_image_f32 & img, clip_gen_process_type gen_process, int n_step, int n_frames)
+        : clip_graph(ctx, img), gen_process(gen_process), n_step(n_step), n_frames(n_frames) {}
+    ggml_cgraph * build() override;
+
+    clip_gen_process_type gen_process;
+    int n_step;   // lsd_decode steps, fixed at graph-build time
+    int n_frames; // GEN_WAV only: number of latents to decode
+
+    // AdaLN modulation: x * (1 + scale) + shift
+    ggml_tensor * modulate(ggml_tensor * x, ggml_tensor * shift, ggml_tensor * scale) const;
+    ggml_tensor * time_embed(const clip_flow_net::time_embd & te, float t) const;
+    ggml_tensor * flow_forward(ggml_tensor * cond, ggml_tensor * x, float s, float t) const;
+};
+
 // one persisted state buffer used by code2wav, see qwen3tts-gen.cpp
 struct c2w_state_slot {
     std::string name;
@@ -326,6 +379,9 @@ struct c2w_state_slot {
 };
 std::vector<c2w_state_slot> list_c2w_state_slots(const clip_hparams & hparams, const clip_model & model);
 
+// same, for the streaming mimi decoder (pocket-tts GEN_WAV)
+std::vector<c2w_state_slot> list_pockettts_state_slots(const clip_hparams & hparams, const clip_model & model);
+
 struct clip_graph_kimik25 : clip_graph {
     clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;

tools/mtmd/models/pockettts-gen.cpp

@@ -0,0 +1,291 @@
+#include "models.h"
+
+#include <cmath>
+
+// pocket-tts generation stages
+//
+// GEN_CODE: backbone hidden state -> next 32-d latent (flow matching) + end-of-speech score
+// GEN_WAV : a window of latents -> PCM, through the mimi decoder
+//
+// there is no codebook anywhere, "codes" in the mtmd API are continuous features here
+
+ggml_tensor * clip_graph_pockettts_gen::modulate(ggml_tensor * x, ggml_tensor * shift, ggml_tensor * scale) const {
+    ggml_tensor * cur = ggml_mul(ctx0, x, ggml_scale_bias(ctx0, scale, 1.0f, 1.0f));
+    return ggml_add(ctx0, cur, shift);
+}
+
+// see TimestepEmbedder in the reference
+ggml_tensor * clip_graph_pockettts_gen::time_embed(const clip_flow_net::time_embd & te, float t) const {
+    // t is a graph-build constant, so the cos/sin table can be folded into a scaled copy
+    ggml_tensor * args = ggml_scale(ctx0, te.freqs, t);
+    ggml_tensor * emb  = ggml_concat(ctx0, ggml_cos(ctx0, args), ggml_sin(ctx0, args), 0);
+
+    ggml_tensor * cur = build_mm(te.up_w, emb);
+    cur = ggml_add(ctx0, cur, te.up_b);
+    cur = ggml_silu(ctx0, cur);
+    cur = build_mm(te.down_w, cur);
+    cur = ggml_add(ctx0, cur, te.down_b);
+
+    // this "RMSNorm" divides by the unbiased variance, not the mean square
+    // it also rescales the input, not the centered value, see _rms_norm() in mlp.py
+    {
+        const int64_t n = cur->ne[0];
+        ggml_tensor * mean = ggml_mean(ctx0, cur);
+        ggml_tensor * dev  = ggml_sub(ctx0, cur, mean);
+        ggml_tensor * var  = ggml_mean(ctx0, ggml_sqr(ctx0, dev));
+        var = ggml_scale_bias(ctx0, var, (float) n / (float) (n - 1), 1e-5f);
+        cur = ggml_div(ctx0, cur, ggml_sqrt(ctx0, var));
+        cur = ggml_mul(ctx0, cur, te.norm);
+    }
+
+    return cur;
+}
+
+// one velocity evaluation: v(cond, s, t, x)
+ggml_tensor * clip_graph_pockettts_gen::flow_forward(ggml_tensor * cond, ggml_tensor * x, float s, float t) const {
+    const auto & flow = model.flow;
+
+    ggml_tensor * cur = build_mm(flow.input_proj_w, x);
+    cur = ggml_add(ctx0, cur, flow.input_proj_b);
+
+    // the two time conditions are averaged, then added to the projected backbone state
+    ggml_tensor * ts = ggml_add(ctx0, time_embed(flow.time[0], s), time_embed(flow.time[1], t));
+    ts = ggml_scale(ctx0, ts, 1.0f / (float) flow.time.size());
+
+    ggml_tensor * c = build_mm(flow.cond_embd_w, cond);
+    c = ggml_add(ctx0, c, flow.cond_embd_b);
+
+    ggml_tensor * y = ggml_add(ctx0, ts, c);
+    cb(y, "flow_cond", -1);
+
+    const int64_t n_ch = flow.blocks.empty() ? 0 : flow.blocks[0].norm_w->ne[0];
+
+    for (size_t il = 0; il < flow.blocks.size(); il++) {
+        const auto & blk = flow.blocks[il];
+
+        ggml_tensor * mod = build_mm(blk.ada_w, ggml_silu(ctx0, y));
+        mod = ggml_add(ctx0, mod, blk.ada_b);
+
+        ggml_tensor * shift = ggml_view_1d(ctx0, mod, n_ch, 0);
+        ggml_tensor * scale = ggml_view_1d(ctx0, mod, n_ch, (size_t) n_ch * mod->nb[0]);
+        ggml_tensor * gate  = ggml_view_1d(ctx0, mod, n_ch, (size_t) 2 * n_ch * mod->nb[0]);
+
+        ggml_tensor * h = build_norm(cur, blk.norm_w, blk.norm_b, NORM_TYPE_NORMAL, 1e-6f, (int) il);
+        h = modulate(h, shift, scale);
+        h = build_mm(blk.up_w, h);
+        h = ggml_add(ctx0, h, blk.up_b);
+        h = ggml_silu(ctx0, h);
+        h = build_mm(blk.down_w, h);
+        h = ggml_add(ctx0, h, blk.down_b);
+
+        cur = ggml_add(ctx0, cur, ggml_mul(ctx0, gate, h));
+        cb(cur, "flow_blk", (int) il);
+    }
+
+    // final layer: the norm has no weights, only the AdaLN modulation
+    ggml_tensor * mod = build_mm(flow.final_ada_w, ggml_silu(ctx0, y));
+    mod = ggml_add(ctx0, mod, flow.final_ada_b);
+
+    ggml_tensor * shift = ggml_view_1d(ctx0, mod, n_ch, 0);
+    ggml_tensor * scale = ggml_view_1d(ctx0, mod, n_ch, (size_t) n_ch * mod->nb[0]);
+
+    cur = build_norm(cur, nullptr, nullptr, NORM_TYPE_NORMAL, 1e-6f, -1);
+    cur = modulate(cur, shift, scale);
+    cur = build_mm(flow.final_proj_w, cur);
+    cur = ggml_add(ctx0, cur, flow.final_proj_b);
+
+    return cur;
+}
+
+// state carried between GEN_WAV calls: rope offset, per-layer KV window, conv left context
+// and the transposed-conv overlap tails
+std::vector<c2w_state_slot> list_pockettts_state_slots(const clip_hparams & hparams, const clip_model & model) {
+    std::vector<c2w_state_slot> slots;
+    if (model.gen_upsample_w == nullptr) {
+        return slots; // not a pocket-tts decoder
+    }
+    const auto & seanet = model.seanet;
+
+    // the slots below are sized from these
+    GGML_ASSERT(!model.gen_tfm_layers.empty());
+    GGML_ASSERT((int) seanet.stages.size() >= hparams.seanet_n_stage);
+    GGML_ASSERT((int) hparams.seanet_ratios.size() >= hparams.seanet_n_stage);
+    GGML_ASSERT(hparams.mimi_tfm_context > 1 && hparams.mimi_downsample > 0);
+
+    slots.push_back({"tfm_pos", 1, 1});
+
+    const int64_t n_embd_a = model.gen_tfm_layers[0].q_w->ne[1];
+    const int64_t prefix   = hparams.mimi_tfm_context - 1;
+    for (size_t il = 0; il < model.gen_tfm_layers.size(); il++) {
+        slots.push_back({"tfm_k_" + std::to_string(il), n_embd_a, prefix});
+        slots.push_back({"tfm_v_" + std::to_string(il), n_embd_a, prefix});
+    }
+
+    // upsample is depthwise, its output channel count is the input one
+    slots.push_back({"up", model.gen_upsample_w->ne[0] - hparams.mimi_downsample, model.gen_upsample_w->ne[2]});
+
+    slots.push_back({"dec_in", seanet.conv_in_w->ne[0] - 1, seanet.conv_in_w->ne[1]});
+    for (int i = 0; i < hparams.seanet_n_stage; i++) {
+        const auto & stage  = seanet.stages[i];
+        const int    stride = hparams.seanet_ratios[hparams.seanet_n_stage - 1 - i];
+        slots.push_back({"dec_up_" + std::to_string(i), stage.scale_conv_w->ne[0] - stride, stage.scale_conv_w->ne[1]});
+        slots.push_back({"dec_res_" + std::to_string(i), stage.res_conv1_w->ne[0] - 1, stage.res_conv1_w->ne[1]});
+    }
+    slots.push_back({"dec_out", seanet.conv_out_w->ne[0] - 1, seanet.conv_out_w->ne[1]});
+
+    return slots;
+}
+
+ggml_cgraph * clip_graph_pockettts_gen::build() {
+    if (gen_process == CLIP_GEN_PROCESS_GEN_CODE) {
+        // the backbone hidden state arrives as the single batch entry
+        ggml_tensor * h_state = build_inp_raw(1);
+        h_state = ggml_reshape_2d(ctx0, h_state, n_mmproj_embd, 1);
+
+        // end-of-speech probe, thresholded on the host side
+        ggml_tensor * eos = build_mm(model.gen_out_eos_w, h_state);
+        eos = ggml_add(ctx0, eos, model.gen_out_eos_b);
+        ggml_set_name(eos, "out_eos_score");
+        ggml_set_output(eos);
+        ggml_build_forward_expand(gf, eos);
+
+        const int64_t n_latent = model.gen_input_lin_w->ne[0];
+
+        ggml_tensor * noise = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_latent, 1);
+        ggml_set_name(noise, "inp_noise");
+        ggml_set_input(noise);
+
+        // lsd_decode: integrate the velocity field from the noise sample
+        ggml_tensor * cur = noise;
+        for (int i = 0; i < n_step; i++) {
+            const float s = (float) i / (float) n_step;
+            const float t = (float) (i + 1) / (float) n_step;
+            ggml_tensor * v = flow_forward(h_state, cur, s, t);
+            cur = ggml_add(ctx0, cur, ggml_scale(ctx0, v, 1.0f / (float) n_step));
+        }
+        cb(cur, "flow_latent", -1);
+
+        ggml_set_name(cur, "out_feats");
+        ggml_set_output(cur);
+        ggml_build_forward_expand(gf, cur);
+
+        // the same latent, projected into the backbone's input space for the next step
+        ggml_tensor * embd = build_mm(model.gen_input_lin_w, cur);
+        cb(embd, "gen_embd", -1);
+        ggml_build_forward_expand(gf, embd);
+
+        return gf;
+    }
+
+    // GEN_WAV: [32, n_frames] latents -> PCM
+    ggml_tensor * feats = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32,
+                                             model.gen_input_lin_w->ne[0], n_frames);
+    ggml_set_name(feats, "inp_feats");
+    ggml_set_input(feats);
+
+    // denormalize, then the DummyQuantizer up-projection
+    ggml_tensor * cur = ggml_add(ctx0, ggml_mul(ctx0, feats, model.gen_emb_std), model.gen_emb_mean);
+    cur = build_mm(model.gen_quant_out_w, cur);
+    cb(cur, "quant_out", -1);
+
+    clip_graph_pockettts_seanet seanet(*this);
+    for (const auto & slot : list_pockettts_state_slots(hparams, model)) {
+        ggml_tensor * t = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, slot.ne0, slot.ne1);
+        ggml_set_name(t, ("state_in_" + slot.name).c_str());
+        ggml_set_input(t);
+        seanet.state_in[slot.name] = t;
+    }
+
+    // model frame rate -> encoder frame rate, depthwise transposed conv
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+    cur = seanet.conv_transpose1d(cur, model.gen_upsample_w, nullptr, hparams.mimi_downsample, "up");
+    cb(cur, "mimi_upsample", -1);
+
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+    // positions continue across calls, the counter lives in the state
+    const int64_t n_pos  = cur->ne[1];
+    const int64_t prefix = hparams.mimi_tfm_context - 1;
+    const int64_t n_kv   = prefix + n_pos;
+
+    ggml_tensor * base    = ggml_reshape_1d(ctx0, seanet.state_in.at("tfm_pos"), 1);
+    ggml_tensor * inp_pos = ggml_cast(ctx0, ggml_add(ctx0, ggml_arange(ctx0, 0.0f, (float) n_pos, 1.0f), base),
+                                      GGML_TYPE_I32);
+    seanet.state_out.push_back({"tfm_pos", ggml_scale_bias(ctx0, seanet.state_in.at("tfm_pos"), 1.0f, (float) n_pos)});
+
+    // banded causal mask over [cached prefix | this chunk]
+    // the last factor masks out cache rows that hold no real frame yet
+    ggml_tensor * pos_k = ggml_reshape_2d(ctx0, ggml_arange(ctx0, 0.0f, (float) n_kv, 1.0f), n_kv, 1);
+    ggml_tensor * pos_q = ggml_reshape_2d(ctx0, ggml_arange(ctx0, (float) prefix, (float) (prefix + n_pos), 1.0f), 1, n_pos);
+    ggml_tensor * diff  = ggml_sub(ctx0, ggml_repeat_4d(ctx0, pos_q, n_kv, n_pos, 1, 1), pos_k);
+
+    ggml_tensor * keep = ggml_mul(ctx0,
+        ggml_step(ctx0, ggml_scale_bias(ctx0, diff, 1.0f, 0.5f)),                                     // delta >= 0
+        ggml_step(ctx0, ggml_scale_bias(ctx0, diff, -1.0f, (float) hparams.mimi_tfm_context - 0.5f))); // delta < context
+    keep = ggml_mul(ctx0, keep,
+        ggml_step(ctx0, ggml_scale_bias(ctx0, ggml_add(ctx0, pos_k, base), 1.0f, 0.5f - (float) prefix)));
+    ggml_tensor * kq_mask = ggml_reshape_4d(ctx0, ggml_log(ctx0, keep), n_kv, n_pos, 1, 1);
+
+    for (int il = 0; il < n_layer; il++) {
+        const auto & layer = model.gen_tfm_layers[il];
+        ggml_tensor * inp = cur;
+
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+
+        ggml_tensor * Qcur = build_mm(layer.q_w, cur);
+        ggml_tensor * Kcur = build_mm(layer.k_w, cur);
+        ggml_tensor * Vcur = build_mm(layer.v_w, cur);
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, d_head, GGML_ROPE_TYPE_NORMAL, 0,
+                             hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, d_head, GGML_ROPE_TYPE_NORMAL, 0,
+                             hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+
+        // prepend the cached window, then keep this chunk's tail for the next call
+        const std::string k_name = "tfm_k_" + std::to_string(il);
+        const std::string v_name = "tfm_v_" + std::to_string(il);
+        ggml_tensor * k_full = ggml_concat(ctx0, seanet.state_in.at(k_name),
+                                           ggml_reshape_2d(ctx0, Kcur, d_head * n_head, n_pos), 1);
+        ggml_tensor * v_full = ggml_concat(ctx0, seanet.state_in.at(v_name), Vcur, 1);
+        seanet.state_out.push_back({k_name, ggml_cont(ctx0, ggml_view_2d(ctx0, k_full, k_full->ne[0], prefix,
+                                                                         k_full->nb[1], (size_t) n_pos * k_full->nb[1]))});
+        seanet.state_out.push_back({v_name, ggml_cont(ctx0, ggml_view_2d(ctx0, v_full, v_full->ne[0], prefix,
+                                                                         v_full->nb[1], (size_t) n_pos * v_full->nb[1]))});
+
+        ggml_tensor * q_cur = ggml_reshape_4d(ctx0, Qcur,   d_head, n_head, n_pos, 1);
+        ggml_tensor * k_cur = ggml_reshape_4d(ctx0, k_full, d_head, n_head, n_kv,  1);
+        ggml_tensor * v_cur = ggml_reshape_4d(ctx0, v_full, d_head, n_head, n_kv,  1);
+
+        cur = build_attn(layer.o_w, nullptr, q_cur, k_cur, v_cur, kq_mask, kq_scale, il);
+        cur = ggml_mul(ctx0, cur, layer.ls_1_w);
+        cur = ggml_add(ctx0, cur, inp);
+
+        inp = cur;
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cur = build_ffn(cur, layer.ff_up_w, nullptr, nullptr, nullptr, layer.ff_down_w, nullptr, FFN_GELU, il);
+        cur = ggml_mul(ctx0, cur, layer.ls_2_w);
+        cur = ggml_add(ctx0, cur, inp);
+    }
+    cb(cur, "mimi_dec_tfm", -1);
+
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+    cur = seanet.decode(cur);
+
+    for (const auto & s : seanet.state_out) {
+        ggml_set_name(s.second, ("state_out_" + s.first).c_str());
+        ggml_set_output(s.second);
+        ggml_build_forward_expand(gf, s.second);
+    }
+
+    // [n_samples, 1] -> [n_samples], clamped like the reference output
+    cur = ggml_reshape_1d(ctx0, cur, cur->ne[0]);
+    cur = ggml_clamp(ctx0, cur, -1.0f, 1.0f);
+    ggml_set_name(cur, "out_audio");
+    ggml_set_output(cur);
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/pockettts-seanet.cpp

@@ -0,0 +1,162 @@
+#include "models.h"
+
+// SEANet convolution stack of the mimi codec, see pocket_tts/modules/seanet.py
+//
+// tensors are T-first here: [T, C]
+// the convs are causal: left context comes from a state slot, or from padding on a cold start
+
+static int64_t div_ceil(int64_t a, int64_t b) {
+    return a / b + (a % b ? 1 : 0);
+}
+
+// x: [T, IC], w: [K, IC, OC] -> [T / stride, OC]
+// the convs are causal, so the whole K - stride padding goes on the left
+ggml_tensor * clip_graph_pockettts_seanet::conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, int dilation,
+                                                  bool pad_replicate, const std::string & state_name) const {
+    const int64_t k_size  = (w->ne[0] - 1) * dilation + 1;
+    const int64_t p_total = k_size - stride;
+
+    // trailing padding so the last frame is not dropped, see pad_for_conv1d() in conv.py
+    const int64_t n_frames  = div_ceil(x->ne[0] - k_size + p_total, stride);
+    const int64_t ideal_len = n_frames * stride + k_size - p_total;
+    const int64_t p_extra   = ideal_len - x->ne[0];
+
+    if (!state_name.empty() && p_total > 0) {
+        // streaming: the left context is the tail of the previous call
+        ggml_tensor * left = state_in.at(state_name); // [p_total, IC]
+        x = ggml_concat(ctx0, left, x, 0);
+        state_out.push_back({state_name,
+            ggml_cont(ctx0, ggml_view_2d(ctx0, x, p_total, x->ne[1], x->nb[1],
+                                         (size_t) (x->ne[0] - p_total) * x->nb[0]))});
+    } else if (pad_replicate && p_total > 0) {
+        // the resamplers repeat the first frame instead of zero-padding
+        ggml_tensor * first = ggml_view_2d(ctx0, x, 1, x->ne[1], x->nb[1], 0);
+        ggml_tensor * left  = ggml_repeat_4d(ctx0, first, p_total, x->ne[1], 1, 1);
+        x = ggml_concat(ctx0, left, x, 0);
+        x = ggml_pad_ext(ctx0, x, 0, p_extra, 0, 0, 0, 0, 0, 0);
+    } else {
+        x = ggml_pad_ext(ctx0, x, p_total, p_extra, 0, 0, 0, 0, 0, 0);
+    }
+
+    ggml_tensor * y = ggml_conv_1d(ctx0, w, x, stride, 0, dilation);
+    y = ggml_reshape_2d(ctx0, y, y->ne[0], y->ne[1]);
+    if (b) {
+        y = ggml_add(ctx0, y, ggml_reshape_2d(ctx0, b, 1, b->ne[0]));
+    }
+    return y;
+}
+
+// x: [T, IC], w: [K, OC/groups, IC] -> [T * stride, OC]
+// the K - stride overlap tail belongs to the next call: added to its head when streaming, else dropped
+ggml_tensor * clip_graph_pockettts_seanet::conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride,
+                                                            const std::string & state_name) const {
+    const int64_t K         = w->ne[0];
+    const int64_t T         = x->ne[0];
+    const int64_t p_total   = K - stride;
+    const bool    depthwise = w->ne[1] == 1 && w->ne[2] > 1;
+    const int64_t OC        = depthwise ? w->ne[2] : w->ne[1];
+    const int64_t emit_len  = T * stride;
+
+    // one column per input step, holding the [K, OC] window that col2im scatter-adds at t * stride
+    ggml_tensor * col;
+    if (depthwise) {
+        // one group per channel: a batched matmul over the channels scales the kernel by each step
+        ggml_tensor * krn = ggml_reshape_3d(ctx0, w, 1, K, OC);             // [1, K, OC]
+        ggml_tensor * xs  = ggml_reshape_3d(ctx0, x, 1, T, OC);             // [1, T, OC]
+        col = ggml_mul_mat(ctx0, krn, xs);                                  // [K, T, OC]
+        col = ggml_cont(ctx0, ggml_permute(ctx0, col, 0, 2, 1, 3));         // [K, OC, T]
+        col = ggml_reshape_2d(ctx0, col, K * OC, T);
+    } else {
+        ggml_tensor * w2 = ggml_reshape_2d(ctx0, w, K * OC, w->ne[2]);
+        w2               = ggml_cont(ctx0, ggml_transpose(ctx0, w2));       // [IC, K * OC]
+        ggml_tensor * xt = ggml_cont(ctx0, ggml_transpose(ctx0, x));        // [IC, T]
+        col = ggml_mul_mat(ctx0, w2, xt);
+    }
+    ggml_tensor * full = ggml_col2im_1d(ctx0, col, stride, OC, 0);          // [emit_len + p_total, OC]
+
+    ggml_tensor * out;
+    if (state_name.empty() || p_total == 0) {
+        out = ggml_cont(ctx0, ggml_view_2d(ctx0, full, emit_len, full->ne[1], full->nb[1], 0));
+    } else {
+        // overlap-add the tail the previous call held back
+        ggml_tensor * prev = state_in.at(state_name); // [p_total, OC]
+        ggml_tensor * head = ggml_add(ctx0, ggml_view_2d(ctx0, full, p_total, full->ne[1], full->nb[1], 0), prev);
+        if (emit_len > p_total) {
+            ggml_tensor * rest = ggml_view_2d(ctx0, full, emit_len - p_total, full->ne[1], full->nb[1],
+                                              (size_t) p_total * full->nb[0]);
+            out = ggml_concat(ctx0, head, rest, 0);
+        } else {
+            out = head;
+        }
+        state_out.push_back({state_name,
+            ggml_cont(ctx0, ggml_view_2d(ctx0, full, p_total, full->ne[1], full->nb[1],
+                                         (size_t) emit_len * full->nb[0]))});
+    }
+
+    if (b) {
+        out = ggml_add(ctx0, out, ggml_reshape_2d(ctx0, b, 1, b->ne[0]));
+    }
+    return out;
+}
+
+ggml_tensor * clip_graph_pockettts_seanet::res_unit(ggml_tensor * x, const clip_seanet::stage & stage, int dilation,
+                                                    const std::string & state_prefix) const {
+    ggml_tensor * h = ggml_elu(ctx0, x);
+    h = conv1d(h, stage.res_conv1_w, stage.res_conv1_b, 1, dilation, false, state_prefix);
+    h = ggml_elu(ctx0, h);
+    // the second conv is pointwise, it needs no left context
+    h = conv1d(h, stage.res_conv2_w, stage.res_conv2_b, 1, 1);
+    return ggml_add(ctx0, x, h);
+}
+
+ggml_tensor * clip_graph_pockettts_seanet::encode(ggml_tensor * x) const {
+    const auto & seanet = model.seanet;
+
+    ggml_tensor * cur = conv1d(x, seanet.conv_in_w, seanet.conv_in_b, 1, 1);
+    cb(cur, "seanet_enc_in", -1);
+
+    for (int i = 0; i < hparams.seanet_n_stage; i++) {
+        const auto & stage  = seanet.stages[i];
+        const int    stride = hparams.seanet_ratios[i];
+
+        cur = res_unit(cur, stage, 1);
+        cur = ggml_elu(ctx0, cur);
+        cur = conv1d(cur, stage.scale_conv_w, stage.scale_conv_b, stride, 1);
+        cb(cur, "seanet_enc_stage", i);
+    }
+
+    cur = ggml_elu(ctx0, cur);
+    cur = conv1d(cur, seanet.conv_out_w, seanet.conv_out_b, 1, 1);
+    cb(cur, "seanet_enc_out", -1);
+
+    return cur;
+}
+
+ggml_tensor * clip_graph_pockettts_seanet::decode(ggml_tensor * x) const {
+    const auto & seanet = model.seanet;
+    const bool   stream = !state_in.empty();
+
+    ggml_tensor * cur = conv1d(x, seanet.conv_in_w, seanet.conv_in_b, 1, 1, false,
+                               stream ? "dec_in" : "");
+    cb(cur, "seanet_dec_in", -1);
+
+    for (int i = 0; i < hparams.seanet_n_stage; i++) {
+        const auto & stage = seanet.stages[i];
+        // the decoder mirrors the encoder, so the ratios are walked backwards
+        const int    stride = hparams.seanet_ratios[hparams.seanet_n_stage - 1 - i];
+        const std::string id = std::to_string(i);
+
+        cur = ggml_elu(ctx0, cur);
+        cur = conv_transpose1d(cur, stage.scale_conv_w, stage.scale_conv_b, stride,
+                               stream ? "dec_up_" + id : "");
+        cur = res_unit(cur, stage, 1, stream ? "dec_res_" + id : "");
+        cb(cur, "seanet_dec_stage", i);
+    }
+
+    cur = ggml_elu(ctx0, cur);
+    cur = conv1d(cur, seanet.conv_out_w, seanet.conv_out_b, 1, 1, false,
+                 stream ? "dec_out" : "");
+    cb(cur, "seanet_dec_out", -1);
+
+    return cur;
+}

tools/mtmd/models/pockettts-spkenc.cpp

@@ -0,0 +1,77 @@
+#include "models.h"
+
+// voice-prompt encoder: raw 24kHz waveform -> one conditioning row per 12.5Hz frame
+// mimi encoder (SEANet + transformer + downsample), then flow_lm.speaker_proj_weight
+
+// pre-norm block with layer scale on both residual paths, see mimi_transformer.py
+ggml_tensor * clip_graph_pockettts_spkenc::tfm_layer_forward(ggml_tensor * cur, const clip_layer & layer, ggml_tensor * inp_pos, ggml_tensor * kq_mask, int il) const {
+    ggml_tensor * inp = cur;
+
+    cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+
+    ggml_tensor * Qcur = build_mm(layer.q_w, cur);
+    ggml_tensor * Kcur = build_mm(layer.k_w, cur);
+    ggml_tensor * Vcur = build_mm(layer.v_w, cur);
+
+    const int64_t n_pos = cur->ne[1];
+    Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
+
+    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, d_head, GGML_ROPE_TYPE_NORMAL, 0,
+                         hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, d_head, GGML_ROPE_TYPE_NORMAL, 0,
+                         hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+
+    cur = build_attn(layer.o_w, nullptr, Qcur, Kcur, Vcur, kq_mask, kq_scale, il);
+    cur = ggml_mul(ctx0, cur, layer.ls_1_w);
+    cur = ggml_add(ctx0, cur, inp);
+
+    inp = cur;
+    cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+    cur = build_ffn(cur, layer.ff_up_w, nullptr, nullptr, nullptr, layer.ff_down_w, nullptr, FFN_GELU, il);
+    cur = ggml_mul(ctx0, cur, layer.ls_2_w);
+    cur = ggml_add(ctx0, cur, inp);
+
+    return cur;
+}
+
+ggml_cgraph * clip_graph_pockettts_spkenc::build() {
+    // the preprocessor hands over the waveform as a single-row "mel", already [n_samples, 1]
+    ggml_tensor * inp_raw = build_inp_raw(1);
+    ggml_tensor * cur = ggml_reshape_2d(ctx0, inp_raw, inp_raw->ne[0], inp_raw->ne[1]);
+
+    clip_graph_pockettts_seanet seanet(*this);
+    cur = seanet.encode(cur);
+    cb(cur, "mimi_enc", -1);
+
+    // [T, 512] -> transformer works on [512, T]
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+    ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cur->ne[1]);
+    ggml_set_name(inp_pos, "inp_pos");
+    ggml_set_input(inp_pos);
+
+    // the mimi transformer is causal with a sliding window, see _build_attention_mask()
+    ggml_tensor * kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, cur->ne[1], cur->ne[1]);
+    ggml_set_name(kq_mask, "kq_mask");
+    ggml_set_input(kq_mask);
+
+    for (int il = 0; il < n_layer; il++) {
+        cur = tfm_layer_forward(cur, model.layers[il], inp_pos, kq_mask, il);
+    }
+    cb(cur, "mimi_enc_tfm", -1);
+
+    // downsample to the model frame rate, [512, T] -> [T, 512] -> [T / 16, 32]
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+    cur = seanet.conv1d(cur, model.downsample_w, nullptr, hparams.mimi_downsample, 1, true);
+    cb(cur, "mimi_downsample", -1);
+
+    // voice latent -> backbone embd
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+    cur = build_mm(model.spk_proj_w, cur);
+    cb(cur, "spk_proj", -1);
+
+    ggml_build_forward_expand(gf, cur);
+    return gf;
+}

tools/mtmd/models/qwen3tts-gen.cpp

@@ -610,6 +610,10 @@ std::vector<c2w_state_slot> list_c2w_state_slots(const clip_hparams & hparams, c
     const auto & c2w = model.c2w;
     std::vector<c2w_state_slot> slots;
 
+    if (c2w.pre_conv_w == nullptr) {
+        return slots; // not a code2wav model, it keeps no state between calls
+    }
+
     slots.push_back({"tfm_pos", 1, 1});
 
     // prefix is (W-1) frames, the batch itself gives the other N=W frames (see tfm_layer_forward)

tools/mtmd/mtmd-audio.cpp

@@ -1423,3 +1423,41 @@ std::vector<float> mtmd_audio_streaming_istft::flush() {
 
     return output;
 }
+
+//
+// mtmd_audio_preprocessor_pockettts
+//
+// mimi takes the raw 24kHz waveform, there is no mel front-end
+// the samples are handed over as a single-row "mel", to reuse the normal chunk path
+//
+
+bool mtmd_audio_preprocessor_pockettts::preprocess(const float *                 samples,
+                                                   size_t                        n_samples,
+                                                   std::vector<mtmd_audio_mel> & output) {
+    // the encoder needs whole frames, see pad_for_conv1d() in the reference
+    const int64_t frame_size = (int64_t) hparams.mimi_downsample * 120;
+    if (n_samples == 0 || frame_size <= 0) {
+        return false;
+    }
+
+    // the mimi transformer mask is dense, so cost is quadratic in the reference length
+    const int64_t max_samples = (int64_t) clip_hparams::pockettts_max_spk_seconds * hparams.audio_sample_rate;
+    if ((int64_t) n_samples > max_samples) {
+        LOG_WRN("%s: speaker reference is %.1f s, truncating to the first %d s\n", __func__,
+                (double) n_samples / hparams.audio_sample_rate, clip_hparams::pockettts_max_spk_seconds);
+        n_samples = (size_t) max_samples;
+    }
+
+    const int64_t n_frames  = (int64_t) (n_samples + frame_size - 1) / frame_size;
+    const int64_t n_padded  = n_frames * frame_size;
+
+    mtmd_audio_mel out;
+    out.n_mel     = 1;
+    out.n_len     = n_padded;
+    out.n_len_org = (int64_t) n_samples;
+    out.data.assign((size_t) n_padded, 0.0f);
+    std::copy(samples, samples + n_samples, out.data.begin());
+
+    output.push_back(std::move(out));
+    return true;
+}

tools/mtmd/mtmd-audio.h

@@ -129,6 +129,13 @@ struct mtmd_audio_preprocessor_qwen3tts_spk : mtmd_audio_preprocessor {
     mtmd_audio_cache cache;
 };
 
+// mimi convolves the waveform directly, so this only pads it to a whole number of frames
+struct mtmd_audio_preprocessor_pockettts : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_pockettts(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
+    void initialize() override {}
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+};
+
 struct mtmd_audio_preprocessor_parakeet : mtmd_audio_preprocessor {
     mtmd_audio_preprocessor_parakeet(clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) { }
     void initialize() override;

tools/mtmd/mtmd-helper-gen.cpp

@@ -5,6 +5,8 @@
 #include "../src/llama-ext.h"
 
 #include <algorithm>
+#include <cctype>
+#include <cmath>
 #include <cstring>
 #include <memory>
 #include <string>
@@ -87,7 +89,8 @@ public:
     virtual int32_t step_prompt(int32_t n_batch) = 0;
     // sampled can be LLAMA_TOKEN_NULL for pipelines with no discrete backbone token,
     // those read what they need from h_state_in instead
-    virtual int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out) = 0;
+    // set out_stop on end-of-speech, h_state_out must be null if no frame is generated
+    virtual int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out, bool * out_stop) = 0;
     virtual int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) = 0;
 
 protected:
@@ -200,8 +203,10 @@ public:
         prompt_pos = 0;
 
         pos = 0;
-        top_k = inp->top_k > 0 ? inp->top_k : 50;
-        top_p = inp->top_p > 0 ? inp->top_p : 1.0f;
+        const mtmd_gen_inp def = mtmd_gen_inp_default(mctx);
+        top_k = inp->top_k > 0 ? inp->top_k : def.top_k;
+        top_p = inp->top_p > 0 ? inp->top_p : def.top_p;
+        seed  = inp->seed;
         out_type = inp->out_type;
 
         // the prompt above holds the whole text stream up to tts_eos, so every generated
@@ -241,13 +246,26 @@ public:
         return n_prompt - prompt_pos;
     }
 
-    int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out) override {
-        mtmd_gen_inp inp{};
+    int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out, bool * out_stop) override {
+        if (sampled == LLAMA_TOKEN_NULL) {
+            LOG_ERR("mtmd_helper_gen_audio: qwen3tts requires a token sampled from the backbone\n");
+            return 1;
+        }
+
+        // backbone signals end-of-speech with a token, no frame for this step
+        if (sampled == codec_eos || llama_vocab_is_eog(vocab, sampled)) {
+            *out_stop    = true;
+            *h_state_out = nullptr;
+            return 0;
+        }
+
+        mtmd_gen_inp inp = mtmd_gen_inp_default(mctx);
         inp.type  = MTMD_GEN_PROCESS_TYPE_GEN_CODE;
         inp.code0 = sampled - codec_0;
         inp.embd  = const_cast<float *>(h_state_in);
         inp.top_k = top_k;
         inp.top_p = top_p;
+        inp.seed  = seed;
         mtmd_gen_out out{};
         if (mtmd_gen_audio_process(mctx, &inp, &out) != 0) {
             LOG_ERR("mtmd_helper_gen_audio: gen_code process failed\n");
@@ -384,10 +402,11 @@ private:
         if (codes_buf.empty()) {
             return true;
         }
-        mtmd_gen_inp inp{};
+        mtmd_gen_inp inp = mtmd_gen_inp_default(mctx);
         inp.type       = MTMD_GEN_PROCESS_TYPE_GEN_WAV;
         inp.codes      = codes_buf.data();
         inp.n_codes    = codes_buf.size();
+        inp.seed       = seed; // same seed as gen_code, else clip reseeds mid-generation
         inp.state_data = c2w_state.empty() ? nullptr : (const char *) c2w_state.data();
         inp.state_size = c2w_state.size();
         mtmd_gen_out out{};
@@ -427,8 +446,9 @@ private:
     std::unique_ptr<decode_embd_batch> prompt_batch;
     int n_prompt = 0;
     int prompt_pos = 0;
-    int32_t top_k = 50;
-    float   top_p = 1.0f;
+    int32_t  top_k = 50;
+    float    top_p = 1.0f;
+    uint32_t seed  = UINT32_MAX;
     std::vector<int32_t> codes_buf;
     std::vector<uint8_t> c2w_state;
     std::vector<float>   audio_pcm;
@@ -438,10 +458,547 @@ private:
     std::vector<char> out_buf;
 };
 
+// settings that only live in the reference's per-pack yaml, not in the checkpoint
+// the english packs share the same shapes and tokenizer, but disagree on these
+// all three are 0 / false when the pack does not tune them, the model default is then used
+struct pockettts_pack_settings {
+    float temp             = 0.0f;
+    int   frames_after_eos = 0;
+    bool  pad_short_text   = false;
+};
+
+static pockettts_pack_settings pockettts_pack(const char * variant) {
+    static const std::unordered_map<std::string, pockettts_pack_settings> packs = {
+        { "english",         { 0.3f, 0, false } },
+        { "english_2026-01", { 0.7f, 0, true  } },
+        { "english_2026-04", { 0.3f, 0, false } },
+        { "french_24l",      { 0.7f, 8, false } },
+    };
+    auto it = packs.find(variant ? variant : "");
+    if (it == packs.end()) {
+        LOG_WRN("mtmd_helper_gen_audio: no tuned settings for pocket-tts variant \"%s\"\n",
+                variant ? variant : "");
+        return {};
+    }
+    return it->second;
+}
+
+// pocket-tts: the backbone emits no token, the flow net turns each hidden state into a latent
+// the end-of-speech head also lives in the mmproj
+class pockettts_gen_audio_pipeline : public mtmd_gen_audio_pipeline {
+public:
+    using mtmd_gen_audio_pipeline::mtmd_gen_audio_pipeline;
+
+    void reset() override {
+        seq_id = 0;
+        pos = 0;
+        feats_buf.clear();
+        dec_state.clear();
+        audio_pcm.clear();
+        h_state_buf.clear();
+        out_buf.clear();
+        prompt_embd_buf.clear();
+        prompt_batch.reset();
+        n_prompt = 0;
+        prompt_pos = 0;
+        step_idx = 0;
+        eos_step = -1;
+        chunks.clear();
+        chunk_idx = 0;
+        n_voice_pos = 0;
+        chunk_budget = 0;
+    }
+
+    int32_t set_input(const mtmd_helper_gen_audio_inp * inp) override {
+        reset();
+        seq_id = inp->seq_id;
+
+        if (!ensure_cache()) {
+            return 1;
+        }
+
+        std::vector<float> voice;
+        if (inp->speaker_ref) {
+            if (!encode_speaker(inp->speaker_ref, voice)) {
+                return 1;
+            }
+        }
+
+        pack = pockettts_pack(info.model_variant);
+
+        const std::string text = prepare_text(std::string(inp->prompt, inp->prompt_len),
+                                              pack.pad_short_text);
+        if (text.empty()) {
+            LOG_ERR("mtmd_helper_gen_audio: empty prompt\n");
+            return 1;
+        }
+
+        std::vector<llama_token> ids(text.size() + 16);
+        int n_ids = llama_tokenize(vocab, text.c_str(), (int32_t) text.size(), ids.data(),
+                                   (int32_t) ids.size(), false, false);
+        if (n_ids <= 0) {
+            LOG_ERR("mtmd_helper_gen_audio: tokenization failed\n");
+            return 1;
+        }
+        ids.resize((size_t) n_ids);
+
+        // long inputs degrade badly, so each chunk restarts from the voice conditioning
+        // see split_into_best_sentences() in the reference
+        chunks = split_chunks(ids);
+        chunk_idx = 0;
+        if (chunks.size() > 1) {
+            LOG_INF("mtmd_helper_gen_audio: %d tokens split into %zu chunks\n", n_ids, chunks.size());
+        }
+
+        const int n_e = n_embd;
+
+        // sequence order is voice, then text, then the audio BOS that starts generation
+        if (!voice.empty()) {
+            GGML_ASSERT(voice.size() % (size_t) n_e == 0);
+            if (bos_before_voice != LLAMA_TOKEN_NULL) {
+                push_embd_row(prompt_embd_buf, bos_before_voice);
+            }
+            prompt_embd_buf.insert(prompt_embd_buf.end(), voice.begin(), voice.end());
+        }
+        // every later chunk rewinds to here and re-prompts, so the voice stays primed
+        n_voice_pos = (int) (prompt_embd_buf.size() / (size_t) n_e);
+
+        for (llama_token t : chunks[0]) {
+            push_embd_row(prompt_embd_buf, t);
+        }
+        push_embd_row(prompt_embd_buf, audio_bos);
+        arm_chunk_budget(0);
+
+        n_prompt = (int) (prompt_embd_buf.size() / (size_t) n_e);
+        prompt_batch.reset(new decode_embd_batch(prompt_embd_buf.data(), n_prompt, 1, n_e));
+        prompt_batch->set_position_normal(0, seq_id);
+        prompt_pos = 0;
+
+        seed     = inp->seed;
+        out_type = inp->out_type;
+
+        return 0;
+    }
+
+    int32_t step_prompt(int32_t n_batch) override {
+        GGML_ASSERT(n_batch > 0);
+        if (prompt_pos >= n_prompt) {
+            return 0;
+        }
+        const int32_t n_tokens_batch = std::min(n_batch, n_prompt - prompt_pos);
+        llama_batch batch_view = prompt_batch->get_view(prompt_pos, n_tokens_batch);
+
+        if ((prompt_pos + n_tokens_batch) == n_prompt) {
+            batch_view.logits[n_tokens_batch - 1] = 1;
+        }
+
+        if (llama_decode(lctx, batch_view) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: prompt decode failed\n");
+            return -1;
+        }
+
+        pos        += n_tokens_batch;
+        prompt_pos += n_tokens_batch;
+
+        if (prompt_pos >= n_prompt) {
+            prompt_batch.reset();
+            prompt_embd_buf.clear();
+            return 0;
+        }
+        return n_prompt - prompt_pos;
+    }
+
+    int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out, bool * out_stop) override {
+        (void) sampled; // the backbone output is continuous, there is no token to consume
+
+        mtmd_gen_inp inp = mtmd_gen_inp_default(mctx);
+        inp.type = MTMD_GEN_PROCESS_TYPE_GEN_CODE;
+        inp.embd = const_cast<float *>(h_state_in);
+        // clip only reseeds when the seed changes, so pass the same one on every step
+        inp.seed = seed;
+        if (pack.temp > 0.0f) {
+            inp.temp = pack.temp;
+        }
+        mtmd_gen_out out{};
+        if (mtmd_gen_audio_process(mctx, &inp, &out) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: flow decode failed\n");
+            return 1;
+        }
+        if (out.is_eos && eos_step < 0) {
+            eos_step = step_idx;
+        }
+        // the frame of the stopping step is discarded, matching _autoregressive_generation().
+        // the budget is the reference's fallback for a chunk whose eos head never fires
+        const bool chunk_done = (eos_step >= 0 && step_idx >= eos_step + frames_after_eos) ||
+                                step_idx >= chunk_budget;
+        if (chunk_done) {
+            if (eos_step < 0) {
+                LOG_WRN("mtmd_helper_gen_audio: chunk %zu hit its budget without end-of-speech\n", chunk_idx);
+            }
+            return finish_chunk(h_state_out, out_stop);
+        }
+
+        feats_buf.insert(feats_buf.end(), out.feats, out.feats + out.n_feats);
+        step_idx++;
+        if (out.n_feats > 0 && feats_buf.size() / out.n_feats >= window_frames) {
+            if (!flush_gen_wav()) {
+                return 1;
+            }
+        }
+
+        decode_embd_batch batch_embd(const_cast<float *>(out.embd), 1, 1, n_embd);
+        batch_embd.set_position_normal(pos, seq_id);
+        batch_embd.batch.logits[0] = 1;
+        pos++;
+
+        if (llama_decode(lctx, batch_embd.batch) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: decode failed\n");
+            return 1;
+        }
+
+        const float * he = llama_get_embeddings_ith(lctx, -1);
+        h_state_buf.assign(he, he + n_embd);
+        *h_state_out = h_state_buf.data();
+
+        return 0;
+    }
+
+    int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) override {
+        if (!flush_gen_wav()) {
+            return 1;
+        }
+
+        *out_sample_rate = info.sample_rate;
+        if (out_n_samples) {
+            *out_n_samples = (int64_t) audio_pcm.size();
+        }
+
+        if (out_type == MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM) {
+            *out_data     = (const char *) audio_pcm.data();
+            *out_data_len = audio_pcm.size() * sizeof(float);
+            return 0;
+        }
+
+        out_buf.clear();
+        if (!write_wav16(out_buf, audio_pcm, info.sample_rate)) {
+            LOG_ERR("mtmd_helper_gen_audio: output too large for WAV\n");
+            return 1;
+        }
+        *out_data     = out_buf.data();
+        *out_data_len = out_buf.size();
+        return 0;
+    }
+
+private:
+    bool ensure_cache() {
+        if (specials_ok) {
+            return true;
+        }
+        // bos_before_voice is optional, some packs do not insert it
+        bos_before_voice = find_special_token(vocab, "<|bos_before_voice|>");
+        audio_bos        = find_special_token(vocab, "<|audio_bos|>");
+        if (audio_bos == LLAMA_TOKEN_NULL) {
+            LOG_ERR("mtmd_helper_gen_audio: missing <|audio_bos|> in vocab\n");
+            return false;
+        }
+        const uint32_t n_tok_embd = llama_model_get_tok_embd(model, nullptr);
+        if (n_tok_embd == 0) {
+            LOG_ERR("mtmd_helper_gen_audio: model has no token embeddings\n");
+            return false;
+        }
+        tok_embd.resize(n_tok_embd);
+        if (llama_model_get_tok_embd(model, tok_embd.data()) != n_tok_embd) {
+            LOG_ERR("mtmd_helper_gen_audio: token embedding copy failed\n");
+            return false;
+        }
+        GGML_ASSERT(n_embd > 0 && n_tok_embd % (uint32_t) n_embd == 0);
+        specials_ok = true;
+        return true;
+    }
+
+    // the table can be shorter than the vocab, so bound the row lookup
+    void push_embd_row(std::vector<float> & dst, llama_token t) const {
+        const size_t n_rows = tok_embd.size() / (size_t) n_embd;
+        GGML_ASSERT(t >= 0 && (size_t) t < n_rows);
+        dst.insert(dst.end(),
+                   tok_embd.begin() + (size_t) t * n_embd,
+                   tok_embd.begin() + (size_t) (t + 1) * n_embd);
+    }
+
+    // token ids of the pieces the reference splits on, see split_into_best_sentences().
+    // the leading token is dropped, it is the tokenizer's dummy prefix
+    std::vector<llama_token> punct_ids(const char * s) const {
+        std::vector<llama_token> ids(16);
+        const int n = llama_tokenize(vocab, s, (int32_t) strlen(s), ids.data(), (int32_t) ids.size(), false, false);
+        if (n <= 1) {
+            return {};
+        }
+        return std::vector<llama_token>(ids.begin() + 1, ids.begin() + n);
+    }
+
+    // cut after runs of boundary tokens, so punctuation stays with the sentence it ends
+    static std::vector<std::vector<llama_token>> split_on(const std::vector<llama_token> & ids,
+                                                          const std::vector<llama_token> & boundary) {
+        std::vector<std::vector<llama_token>> out;
+        size_t start = 0;
+        bool prev_was_boundary = false;
+        for (size_t i = 0; i < ids.size(); i++) {
+            const bool is_boundary = std::find(boundary.begin(), boundary.end(), ids[i]) != boundary.end();
+            if (!is_boundary && prev_was_boundary) {
+                out.emplace_back(ids.begin() + start, ids.begin() + i);
+                start = i;
+            }
+            prev_was_boundary = is_boundary;
+        }
+        out.emplace_back(ids.begin() + start, ids.end());
+        return out;
+    }
+
+    std::vector<std::vector<llama_token>> split_chunks(const std::vector<llama_token> & ids) const {
+        if ((int) ids.size() <= max_chunk_tokens) {
+            return { ids };
+        }
+        const std::vector<llama_token> eos_punct = punct_ids(".!...?");
+        const std::vector<llama_token> mid_punct = punct_ids(",;:");
+
+        // oversized sentences are split again on weaker punctuation, else words get skipped
+        std::vector<std::vector<llama_token>> segments;
+        for (auto & seg : split_on(ids, eos_punct)) {
+            if ((int) seg.size() <= max_chunk_tokens) {
+                segments.push_back(std::move(seg));
+                continue;
+            }
+            auto sub = split_on(seg, mid_punct);
+            if (sub.size() > 1) {
+                for (auto & s : sub) {
+                    segments.push_back(std::move(s));
+                }
+            } else {
+                segments.push_back(std::move(seg));
+            }
+        }
+
+        std::vector<std::vector<llama_token>> out;
+        for (auto & seg : segments) {
+            if (seg.empty()) {
+                continue;
+            }
+            if (!out.empty() && (int) (out.back().size() + seg.size()) <= max_chunk_tokens) {
+                out.back().insert(out.back().end(), seg.begin(), seg.end());
+            } else {
+                out.push_back(std::move(seg));
+            }
+        }
+        if (out.empty()) {
+            out.push_back(ids);
+        }
+        for (const auto & c : out) {
+            if ((int) c.size() > max_chunk_tokens) {
+                LOG_WRN("mtmd_helper_gen_audio: chunk of %zu tokens exceeds the %d token budget, "
+                        "generation may skip words\n", c.size(), max_chunk_tokens);
+            }
+        }
+        return out;
+    }
+
+    // _estimate_max_gen_len() plus the per-chunk tail guess, both in frames
+    void arm_chunk_budget(size_t idx) {
+        const int n_tok = (int) chunks[idx].size();
+        chunk_budget = (int) std::ceil((n_tok / 3.0 + 2.0) * frame_rate);
+        // the pack may pin the tail, else the reference guesses it from the word count
+        frames_after_eos = pack.frames_after_eos > 0 ? pack.frames_after_eos : (n_tok <= 6 ? 5 : 3);
+        step_idx = 0;
+        eos_step = -1;
+    }
+
+    // ends the current chunk and, if there is another, re-prompts it on top of the voice
+    int32_t finish_chunk(const float ** h_state_out, bool * out_stop) {
+        if (!flush_gen_wav()) {
+            return 1;
+        }
+        // the decoder restarts too, the next chunk's audio is not continuous with this one
+        dec_state.clear();
+
+        if (chunk_idx + 1 >= chunks.size()) {
+            *out_stop    = true;
+            *h_state_out = nullptr;
+            return 0;
+        }
+        chunk_idx++;
+
+        // drop this chunk's text and audio, keep the voice conditioning
+        llama_memory_seq_rm(llama_get_memory(lctx), seq_id, n_voice_pos, -1);
+        pos = n_voice_pos;
+
+        const int n_e = n_embd;
+        prompt_embd_buf.clear();
+        for (llama_token t : chunks[chunk_idx]) {
+            push_embd_row(prompt_embd_buf, t);
+        }
+        push_embd_row(prompt_embd_buf, audio_bos);
+        arm_chunk_budget(chunk_idx);
+
+        const int n_rows = (int) (prompt_embd_buf.size() / (size_t) n_e);
+        GGML_ASSERT(n_rows > 0);
+        decode_embd_batch batch(prompt_embd_buf.data(), n_rows, 1, n_e);
+        batch.set_position_normal(pos, seq_id);
+        batch.batch.logits[n_rows - 1] = 1;
+        if (llama_decode(lctx, batch.batch) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: chunk prompt decode failed\n");
+            return 1;
+        }
+        pos += n_rows;
+        prompt_embd_buf.clear();
+
+        const float * he = llama_get_embeddings_ith(lctx, -1);
+        h_state_buf.assign(he, he + n_embd);
+        *h_state_out = h_state_buf.data();
+        *out_stop    = false;
+        return 0;
+    }
+
+    // same normalization as prepare_text_prompt() in the reference, it affects quality
+    static std::string prepare_text(const std::string & in, bool pad_short) {
+        std::string s;
+        s.reserve(in.size() + 1);
+        for (char c : in) {
+            if (c == '\n' || c == '\r') {
+                s += ' ';
+            } else if (c == ';') {
+                s += ',';
+            } else {
+                s += c;
+            }
+        }
+        const size_t b = s.find_first_not_of(' ');
+        const size_t e = s.find_last_not_of(' ');
+        if (b == std::string::npos) {
+            return "";
+        }
+        s = s.substr(b, e - b + 1);
+        if (s[0] >= 'a' && s[0] <= 'z') {
+            s[0] = (char) (s[0] - 'a' + 'A');
+        }
+        const unsigned char last = (unsigned char) s.back();
+        if (std::isalnum(last)) {
+            s += '.';
+        }
+        if (pad_short && count_words(s) < 5) {
+            s = std::string(8, ' ') + s;
+        }
+        return s;
+    }
+
+    static int count_words(const std::string & s) {
+        int n = 0;
+        bool in_word = false;
+        for (char c : s) {
+            if (c == ' ') {
+                in_word = false;
+            } else if (!in_word) {
+                in_word = true;
+                n++;
+            }
+        }
+        return n;
+    }
+
+    // runs the reference wav through the mimi encoder, returns one row per 12.5Hz frame
+    bool encode_speaker(mtmd_bitmap * bitmap, std::vector<float> & out) {
+        if (!mtmd_support_audio(mctx)) {
+            LOG_ERR("mtmd_helper_gen_audio: mmproj has no voice encoder\n");
+            return false;
+        }
+        const std::string  marker = mtmd_default_marker();
+        mtmd_input_text     text{ marker.c_str(), marker.size(), false, true };
+        mtmd_input_chunks * chunks = mtmd_input_chunks_init();
+        const mtmd_bitmap * bptr = bitmap;
+        bool ok = mtmd_tokenize(mctx, chunks, &text, &bptr, 1) == 0;
+        if (ok) {
+            ok = false;
+            for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+                const mtmd_input_chunk * chunk = mtmd_input_chunks_get(chunks, i);
+                if (mtmd_input_chunk_get_type(chunk) != MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+                    continue;
+                }
+                if (mtmd_encode_chunk(mctx, chunk) != 0) {
+                    LOG_ERR("mtmd_helper_gen_audio: voice encode failed\n");
+                    break;
+                }
+                const float * embd = mtmd_get_output_embd(mctx);
+                const size_t  n = (size_t) llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk);
+                out.assign(embd, embd + n);
+                ok = true;
+                break;
+            }
+        }
+        mtmd_input_chunks_free(chunks);
+        return ok;
+    }
+
+    // decodes the buffered latents, the mimi decoder state carries over between calls
+    bool flush_gen_wav() {
+        if (feats_buf.empty()) {
+            return true;
+        }
+        mtmd_gen_inp inp = mtmd_gen_inp_default(mctx);
+        inp.type       = MTMD_GEN_PROCESS_TYPE_GEN_WAV;
+        inp.feats      = feats_buf.data();
+        inp.n_feats    = feats_buf.size();
+        inp.seed       = seed;
+        inp.state_data = dec_state.empty() ? nullptr : (const char *) dec_state.data();
+        inp.state_size = dec_state.size();
+        mtmd_gen_out out{};
+        if (mtmd_gen_audio_process(mctx, &inp, &out) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: mimi decode failed\n");
+            return false;
+        }
+        audio_pcm.insert(audio_pcm.end(), out.audio, out.audio + out.n_samples);
+        dec_state.assign(out.state_data, out.state_data + out.state_size);
+        feats_buf.clear();
+        return true;
+    }
+
+    pockettts_pack_settings pack;
+    bool specials_ok = false;
+    llama_token bos_before_voice = LLAMA_TOKEN_NULL;
+    llama_token audio_bos        = LLAMA_TOKEN_NULL;
+    std::vector<float> tok_embd;
+
+    llama_seq_id seq_id = 0;
+    int pos = 0;
+    std::vector<float> prompt_embd_buf;
+    std::unique_ptr<decode_embd_batch> prompt_batch;
+    int n_prompt = 0;
+    int prompt_pos = 0;
+    uint32_t seed = UINT32_MAX;
+    // end-of-speech is latched, then a few more frames are generated as tail padding
+    int step_idx = 0;
+    int eos_step = -1;
+    int frames_after_eos = 3;
+    static constexpr int max_chunk_tokens = 50;  // MAX_TOKEN_PER_CHUNK in the reference
+    static constexpr double frame_rate    = 12.5;
+    std::vector<std::vector<llama_token>> chunks;
+    size_t chunk_idx   = 0;
+    int    n_voice_pos = 0; // KV positions held by the voice conditioning
+    int    chunk_budget = 0;
+
+    // latents are decoded a window at a time, the decoder state bridges the windows
+    size_t window_frames = 8;
+    std::vector<float>   feats_buf;
+    std::vector<uint8_t> dec_state;
+    std::vector<float>   audio_pcm;
+    std::vector<float>   h_state_buf;
+    mtmd_helper_gen_audio_outtype out_type = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
+    std::vector<char> out_buf;
+};
+
 static std::unique_ptr<mtmd_gen_audio_pipeline> make_pipeline(llama_context * lctx, mtmd_context * mctx) {
     switch (mtmd_gen_audio_get_info(mctx).type) {
         case MTMD_GEN_AUDIO_TYPE_QWEN3TTS:
             return std::unique_ptr<mtmd_gen_audio_pipeline>(new qwen3tts_gen_audio_pipeline(lctx, mctx));
+        case MTMD_GEN_AUDIO_TYPE_POCKETTTS:
+            return std::unique_ptr<mtmd_gen_audio_pipeline>(new pockettts_gen_audio_pipeline(lctx, mctx));
         default:
             return nullptr;
     }
@@ -483,11 +1040,17 @@ int32_t mtmd_helper_gen_audio_step_prompt(mtmd_helper_gen_audio * ctx, int32_t n
 }
 
 int32_t mtmd_helper_gen_audio_step_gen(mtmd_helper_gen_audio * ctx, llama_token sampled,
-                                       const float * h_state_in, const float ** h_state_out) {
+                                       const float * h_state_in, const float ** h_state_out,
+                                       bool * out_stop) {
     if (!ctx->pipeline) {
         return 1;
     }
-    return ctx->pipeline->step_gen(sampled, h_state_in, h_state_out);
+    bool stop = false;
+    const int32_t ret = ctx->pipeline->step_gen(sampled, h_state_in, h_state_out, &stop);
+    if (out_stop) {
+        *out_stop = stop;
+    }
+    return ret;
 }
 
 int32_t mtmd_helper_gen_audio_get_output(mtmd_helper_gen_audio * ctx, int32_t * out_sample_rate,

tools/mtmd/mtmd-helper.h

@@ -183,8 +183,9 @@ struct mtmd_helper_gen_audio_inp {
     mtmd_bitmap * speaker_ref; // optional, can be NULL
     const char * lang; // optional, can be NULL
 
-    int32_t top_k;
-    float   top_p;
+    int32_t  top_k;
+    float    top_p;
+    uint32_t seed; // UINT32_MAX for random (default: random)
 
     enum mtmd_helper_gen_audio_outtype out_type;
 };
@@ -208,12 +209,15 @@ MTMD_API int32_t mtmd_helper_gen_audio_step_prompt(
                         int32_t n_batch);
 
 // generates one frame; must only be called after step_prompt() has returned 0
-// h_state_out is valid until next step_gen() or reset() call
+// sampled can be LLAMA_TOKEN_NULL for pipelines with no discrete backbone token
+// out_stop (optional) is set on end-of-speech, the caller must then stop the loop
+// h_state_out is valid until next step_gen() or reset() call, null if no frame is generated
 MTMD_API int32_t mtmd_helper_gen_audio_step_gen(
                         mtmd_helper_gen_audio * ctx,
                         llama_token sampled,
                         const float *  h_state_in,
-                        const float ** h_state_out);
+                        const float ** h_state_out,
+                        bool * out_stop);
 
 // out_data valid until next get_output() or reset() call
 // out_n_samples (optional, can be NULL) receives the number of generated PCM samples
@@ -261,8 +265,8 @@ struct gen_audio {
     int32_t step_prompt(int32_t n_batch) {
         return mtmd_helper_gen_audio_step_prompt(ctx.get(), n_batch);
     }
-    int32_t step_gen(llama_token sampled, const float * h_state, const float ** h_state_out) {
-        return mtmd_helper_gen_audio_step_gen(ctx.get(), sampled, h_state, h_state_out);
+    int32_t step_gen(llama_token sampled, const float * h_state, const float ** h_state_out, bool * out_stop = nullptr) {
+        return mtmd_helper_gen_audio_step_gen(ctx.get(), sampled, h_state, h_state_out, out_stop);
     }
     int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples = nullptr) {
         return mtmd_helper_gen_audio_get_output(ctx.get(), out_sample_rate, out_data, out_data_len, out_n_samples);

tools/mtmd/mtmd.cpp

@@ -477,6 +477,7 @@ struct mtmd_context {
     // generation context
     struct clip_ctx * ctx_gen_a; // audio
     std::vector<int32_t> gen_out_codes; // this frame's 16 sampled codes (GEN_CODE)
+    std::vector<float>   gen_out_feats; // this frame's continuous features, if any (GEN_CODE)
     std::vector<float>   gen_out_embd;  // next-step hidden state fed back to backbone (GEN_CODE)
     std::vector<float>   gen_out_audio; // decoded PCM samples for the current frame (GEN_WAV)
     std::vector<uint8_t> gen_out_state; // state to feed into the next GEN_WAV call
@@ -979,6 +980,10 @@ struct mtmd_context {
                 {
                     audio_preproc = std::make_unique<mtmd_audio_preprocessor_qwen3tts_spk>(ctx_a);
                 } break;
+            case PROJECTOR_TYPE_POCKETTTS_SPKENC:
+                {
+                    audio_preproc = std::make_unique<mtmd_audio_preprocessor_pockettts>(ctx_a);
+                } break;
             default:
                 throw std::runtime_error(string_format("%s: unexpected audio projector type %d\n", __func__, proj));
         }
@@ -1798,16 +1803,22 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
 //
 
 mtmd_gen_audio_info mtmd_gen_audio_get_info(const mtmd_context * ctx) {
-    mtmd_gen_audio_info info;
+    mtmd_gen_audio_info info{};
+    info.model_variant = "";
     if (!ctx->ctx_gen_a) {
         info.type = MTMD_GEN_AUDIO_TYPE_NONE;
         return info;
     }
+    info.model_variant = clip_get_hparams(ctx->ctx_gen_a)->gen_model_variant.c_str();
     switch (clip_get_projector_type(ctx->ctx_gen_a)) {
         case PROJECTOR_TYPE_QWEN3TTS_GEN:
             info.type = MTMD_GEN_AUDIO_TYPE_QWEN3TTS;
             info.sample_rate = 24000;
             break;
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            info.type = MTMD_GEN_AUDIO_TYPE_POCKETTTS;
+            info.sample_rate = 24000;
+            break;
         default:
             info.type = MTMD_GEN_AUDIO_TYPE_NONE;
             break;
@@ -1815,6 +1826,33 @@ mtmd_gen_audio_info mtmd_gen_audio_get_info(const mtmd_context * ctx) {
     return info;
 }
 
+mtmd_gen_inp mtmd_gen_inp_default(const mtmd_context * ctx) {
+    mtmd_gen_inp inp{};
+    inp.type = MTMD_GEN_PROCESS_TYPE_GEN_CODE;
+    inp.seed = UINT32_MAX;
+    if (!ctx->ctx_gen_a) {
+        return inp;
+    }
+
+    switch (clip_get_projector_type(ctx->ctx_gen_a)) {
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            // https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/blob/main/generation_config.json
+            inp.top_k = 50;
+            inp.top_p = 1.0f;
+            inp.temp  = 0.9f; // TODO: handle this on graph
+            break;
+        case PROJECTOR_TYPE_POCKETTTS_GEN:
+            // https://github.com/kyutai-labs/pocket-tts/blob/main/pocket_tts/default_parameters.py
+            inp.top_k = 50;
+            inp.top_p = 1.0f;
+            inp.temp  = 0.7f;
+            break;
+        default:
+            break;
+    }
+    return inp;
+}
+
 static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_inp * inp, mtmd_gen_out * out) {
     clip_ctx * ctx_clip = ctx->ctx_gen_a;
     if (!ctx_clip) {
@@ -1822,6 +1860,8 @@ static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_in
         return 1;
     }
 
+    *out = {};
+
     if (inp->type == MTMD_GEN_PROCESS_TYPE_GEN_CODE) {
         const size_t n_embd = (size_t) clip_n_mmproj_embd(ctx_clip);
 
@@ -1835,16 +1875,22 @@ static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_in
 
         std::vector<float>   out_embd(n_embd);
         std::vector<int32_t> out_codes;
+        std::vector<float>   out_feats;
+        bool is_eos = false;
 
         clip_encode_params params;
-        params.imgs        = &batch;
-        params.n_threads   = ctx->n_threads;
-        params.gen_process = CLIP_GEN_PROCESS_GEN_CODE;
-        params.out_embd    = &out_embd;
-        params.out_codes   = &out_codes;
-        params.code0       = inp->code0;
-        params.top_k       = inp->top_k;
-        params.top_p       = inp->top_p;
+        params.imgs          = &batch;
+        params.n_threads     = ctx->n_threads;
+        params.gen_process   = CLIP_GEN_PROCESS_GEN_CODE;
+        params.out_embd      = &out_embd;
+        params.out_codes     = &out_codes;
+        params.out_feats     = &out_feats;
+        params.code0         = inp->code0;
+        params.top_k         = inp->top_k;
+        params.top_p         = inp->top_p;
+        params.seed          = inp->seed;
+        params.temp          = inp->temp;
+        params.out_is_eos    = &is_eos;
 
         if (!clip_encode(ctx_clip, &params)) {
             LOG_ERR("%s: clip_encode failed (gen_code)\n", __func__);
@@ -1853,19 +1899,31 @@ static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_in
 
         ctx->gen_out_embd  = std::move(out_embd);
         ctx->gen_out_codes = std::move(out_codes);
+        ctx->gen_out_feats = std::move(out_feats);
 
-        out->embd    = ctx->gen_out_embd.data();
-        out->codes   = ctx->gen_out_codes.data();
-        out->n_codes = ctx->gen_out_codes.size();
+        out->embd      = ctx->gen_out_embd.data();
+        out->codes     = ctx->gen_out_codes.data();
+        out->n_codes   = ctx->gen_out_codes.size();
+        out->feats     = ctx->gen_out_feats.data();
+        out->n_feats   = ctx->gen_out_feats.size();
+        out->is_eos    = is_eos;
         return 0;
     }
 
     // MTMD_GEN_PROCESS_TYPE_GEN_WAV
-    if (!inp->codes || inp->n_codes == 0) {
-        LOG_ERR("%s: codes required for gen_wav\n", __func__);
+    const bool has_codes = inp->codes && inp->n_codes > 0;
+    const bool has_feats = inp->feats && inp->n_feats > 0;
+    if (has_codes == has_feats) {
+        LOG_ERR("%s: gen_wav requires exactly one of codes or feats\n", __func__);
         return 1;
     }
-    std::vector<int32_t> in_codes(inp->codes, inp->codes + inp->n_codes);
+    std::vector<int32_t> in_codes;
+    std::vector<float>   in_feats;
+    if (has_codes) {
+        in_codes.assign(inp->codes, inp->codes + inp->n_codes);
+    } else {
+        in_feats.assign(inp->feats, inp->feats + inp->n_feats);
+    }
     std::vector<uint8_t> in_state;
     if (inp->state_data) {
         in_state.assign(inp->state_data, inp->state_data + inp->state_size);
@@ -1885,7 +1943,10 @@ static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_in
     params.imgs        = &batch;
     params.n_threads   = ctx->n_threads;
     params.gen_process = CLIP_GEN_PROCESS_GEN_WAV;
-    params.codes       = &in_codes;
+    // gen_wav draws no randomness, but keep the seed so it does not reseed mid-generation
+    params.seed        = inp->seed;
+    params.codes       = has_codes ? &in_codes : nullptr;
+    params.feats       = has_feats ? &in_feats : nullptr;
     params.out_audio   = &ctx->gen_out_audio;
     params.state_in    = inp->state_data ? &in_state : nullptr;
     params.state_out   = &ctx->gen_out_state;

tools/mtmd/mtmd.h

@@ -344,18 +344,25 @@ MTMD_API struct mtmd_caps mtmd_get_cap_from_file(const char * mmproj_fname);
 enum mtmd_gen_audio_type {
     MTMD_GEN_AUDIO_TYPE_NONE, // not supported
     MTMD_GEN_AUDIO_TYPE_QWEN3TTS,
+    MTMD_GEN_AUDIO_TYPE_POCKETTTS,
 };
+
 struct mtmd_gen_audio_info {
     enum mtmd_gen_audio_type type;
     int32_t sample_rate; // in Hz, for example 24000 for qwen3tts
+    const char * model_variant; // name of the weight variant, can be nullptr if not applicable
 };
+
 MTMD_API struct mtmd_gen_audio_info mtmd_gen_audio_get_info(const mtmd_context * ctx);
 
+
 enum mtmd_gen_process_type {
     MTMD_GEN_PROCESS_TYPE_GEN_CODE, // h_state to semantic (codes, mel-spectrogram, etc.)
     MTMD_GEN_PROCESS_TYPE_GEN_WAV,  // convert semantic to PCM audio
                                     // for qwen3tts, this is code2wav
+                                    // for pocket-tts, this is mimi decoder
 };
+
 struct mtmd_gen_inp {
     enum mtmd_gen_process_type type;
 
@@ -364,21 +371,30 @@ struct mtmd_gen_inp {
     float * embd;   // the hidden state from backbone, must have n_text_embd elements
     int32_t top_k;
     float   top_p;
+    uint32_t seed; // UINT32_MAX for random
+    float    temp; // sampling temperature, or noise scale for flow-matching decoders
 
     // for MTMD_GEN_PROCESS_TYPE_GEN_WAV
+    // pass either codes (discrete) or feats (continuous), depending on the pipeline
     int32_t * codes;
     size_t    n_codes;
+    const float * feats;
+    size_t        n_feats;
     const char * state_data;
     size_t       state_size;
 };
+
 struct mtmd_gen_out {
     // note: output memory is allocated by the context, valid until next process() call
 
     // for MTMD_GEN_PROCESS_TYPE_GEN_CODE
     const int32_t * codes;
-    size_t n_codes;
+    size_t          n_codes;
+    const float * feats; // continuous counterpart of codes
+    size_t        n_feats;
     const float * embd; // the generated hidden state, to be fed back to backbone
                         // it must have n_text_embd elements
+    bool is_eos; // only set by pipelines having the EOS head inside mmproj
 
     // for MTMD_GEN_PROCESS_TYPE_GEN_WAV
     const float * audio;
@@ -386,6 +402,10 @@ struct mtmd_gen_out {
     const char * state_data;
     size_t       state_size;
 };
+
+// defaults tuned for the loaded pipeline, callers override only what they care about
+MTMD_API struct mtmd_gen_inp mtmd_gen_inp_default(const mtmd_context * ctx);
+
 // note: this API is stateless, caller must handle state management and audio frame accumulation
 MTMD_API int32_t mtmd_gen_audio_process(mtmd_context * ctx,
                                 const struct mtmd_gen_inp * inp,

tools/mtmd/requirements.txt

@@ -2,11 +2,5 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=11.3.0
 
-## Embedding Gemma requires PyTorch 2.6.0 or later, bumped to 2.11.0 for compatibility
-torch==2.11.0; platform_machine != "s390x" # check_requirements: ignore "=="
+torch==2.11.0 # check_requirements: ignore "=="
 torchvision==0.26.0; platform_machine != "s390x" # check_requirements: ignore "=="
-
-# torch s390x packages can only be found from nightly builds
---extra-index-url https://download.pytorch.org/whl/nightly
-torch>=0.0.0.dev0; platform_machine == "s390x" # check_requirements: ignore "=="
-torchvision>=0.0.0.dev0; platform_machine == "s390x" # check_requirements: ignore "=="

tools/server/README.md

@@ -75,7 +75,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |

tools/server/server-context.cpp

@@ -397,12 +397,7 @@ struct server_slot {
 
     bool need_embd() const {
         GGML_ASSERT(task);
-        return task->need_embd() || (spec && common_speculative_need_embd(spec));
-    }
-
-    bool need_embd_nextn() const {
-        GGML_ASSERT(task);
-        return spec && common_speculative_need_embd_nextn(spec);
+        return task->need_embd();
     }
 
     // if the context does not have a memory module then all embeddings have to be computed within a single ubatch

tools/server/tests/tests.sh

@@ -6,11 +6,6 @@ cd $SCRIPT_DIR
 
 set -eu
 
-if [[ "${SLOW_TESTS:-0}" == 1 ]]; then
-    # Slow tests for tool calls need quite a few models ahead of time to avoid timing out.
-    python $SCRIPT_DIR/../../../scripts/fetch_server_test_models.py
-fi
-
 if [ $# -lt 1 ]
 then
     if [[ "${SLOW_TESTS:-0}" == 1 ]]; then

tools/tts/README.md

@@ -32,3 +32,28 @@ llama-tts -hf ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF \
     --tts-speaker-file speaker.mp3 \
     --output out.wav
 ```
+
+## Pocket TTS
+
+Available params:
+- `--tts-speaker-file` should point to a speaker reference audio file (wav, mp3). It is required, the model produces almost no audio without it
+- Note: `lang` is not used, the language is a property of the weights
+
+Example usage:
+
+```sh
+llama-tts -m pocket-tts.gguf \
+    -mm mmproj-pocket-tts.gguf \
+    -p "Hello world" \
+    --tts-speaker-file speaker.mp3 \
+    --output out.wav
+```
+
+**Note for GGUF conversion:**
+
+The [upstream repository](https://huggingface.co/kyutai/pocket-tts) holds one complete model per language under `languages/`, next to a set of shared files at the root. Convert one of the `languages/<name>` directories, **not** the root directory:
+
+```sh
+python convert_hf_to_gguf.py path/to/pocket-tts/languages/english --outfile pocket-tts.gguf
+python convert_hf_to_gguf.py path/to/pocket-tts/languages/english --mmproj --outfile mmproj-pocket-tts.gguf
+```

tools/tts/tts.cpp

@@ -119,6 +119,7 @@ int main(int argc, char ** argv) {
     inp.lang        = params.tts_lang.c_str();
     inp.top_k       = params.sampling.top_k;
     inp.top_p       = params.sampling.top_p;
+    inp.seed        = params.sampling.seed;
     inp.out_type    = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
 
     //
@@ -143,8 +144,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
+    // note: some pipelines ignore this token and use the hidden state instead
     auto sample_semantic_code = [&]() -> llama_token {
         llama_token t = common_sampler_sample(smpl, lctx, -1);
         common_sampler_accept(smpl, t, true);
@@ -159,19 +159,24 @@ int main(int argc, char ** argv) {
     tts_timings timings;
     const int64_t t_gen_start_us = ggml_time_us();
 
-    for (; n_frames < max_new && !llama_vocab_is_eog(vocab, sampled); n_frames++) {
+    bool stop = false;
+    while (!stop && n_frames < max_new) {
         const float * h_next = nullptr;
 
         // stage 2+3: semantic --> acoustic details --> audio waveform
         //            step_gen() runs both stages and returns new h_state for next step
-        if (gen.step_gen(sampled, h_state, &h_next) != 0) {
+        if (gen.step_gen(sampled, h_state, &h_next, &stop) != 0) {
             LOG_ERR("step_gen failed at frame %d\n", n_frames);
             return 1;
         }
+        if (!h_next) {
+            break; // stopped without generating a frame
+        }
 
+        n_frames++;
         h_state = h_next;
         sampled = sample_semantic_code();
-        timings.report(n_frames + 1);
+        timings.report(n_frames);
     }
     const double t_gen_s = (ggml_time_us() - t_gen_start_us) / 1e6;
 

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeDetails.svelte

@@ -3,6 +3,7 @@
 	import { ChevronDown } from '@lucide/svelte';
 	import * as Collapsible from '$lib/components/ui/collapsible';
 	import { STATS_UNITS } from '$lib/constants';
+	import { gaugePopup } from '$lib/stores/context-gauge-popup.svelte';
 
 	interface Props {
 		currentRead: number;
@@ -30,19 +31,19 @@
 		transientDetails
 	}: Props = $props();
 
-	let open = $state(false);
-
 	const hasCumulative = $derived(cumulativeRead > 0 || cumulativeOutput > 0);
 	const hasCurrent = $derived(currentRead > 0 || currentOutput > 0);
 </script>
 
-<Collapsible.Root bind:open class="mt-3 border-t border-border/50 pt-4">
+<Collapsible.Root bind:open={gaugePopup.detailsOpen} class="mt-3 border-t border-border/50 pt-4">
 	<Collapsible.Trigger
 		class="flex w-full cursor-pointer items-center gap-1 text-xs text-muted-foreground hover:text-foreground"
 	>
 		<span>Token usage details</span>
 
-		<ChevronDown class={'ml-auto h-3 w-3 transition-transform' + (open ? ' rotate-180' : '')} />
+		<ChevronDown
+			class={'ml-auto h-3 w-3 transition-transform' + (gaugePopup.detailsOpen ? ' rotate-180' : '')}
+		/>
 	</Collapsible.Trigger>
 
 	<Collapsible.Content class="flex flex-col gap-4 text-xs pt-4">

tools/ui/src/lib/hooks/use-context-gauge.svelte.ts

@@ -112,7 +112,7 @@ export function useContextGauge(): UseContextGaugeReturn {
 		return chatStore.getConversationModel(activeMessages() as DatabaseMessage[]);
 	});
 	const isActiveModelLoaded = $derived(
-		activeModelId !== null && modelsStore.isModelLoaded(activeModelId)
+		activeModelId !== null && (!isRouterMode() || modelsStore.isModelLoaded(activeModelId))
 	);
 	const isActiveModelLoading = $derived(
 		activeModelId !== null && modelsStore.isModelOperationInProgress(activeModelId)

tools/ui/src/lib/stores/context-gauge-popup.svelte.ts

@@ -16,7 +16,7 @@ import {
 let closeTimer: ReturnType<typeof setTimeout> | undefined;
 let lastPointerType = '';
 
-export const gaugePopup = $state({ bottom: 0, centerX: 0, open: false });
+export const gaugePopup = $state({ bottom: 0, centerX: 0, detailsOpen: false, open: false });
 
 function openFrom(trigger: HTMLElement): void {
 	clearTimeout(closeTimer);