mtmd: support pocket-tts (#26871)

* adapt the api

* text model ok

* working impl, need verify and clean up

* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im

ggml_conv_transpose_1d has no grouped mode, so the depthwise upsample
was built as one convolution and one concat per channel, which floods
the graph with small nodes and makes kernel launches dominate the
decoder.

Fold both cases into the column form the seanet decoder already needs:
the general case reshapes the kernel to [IC, K * OC] and matmuls it
with the input, the depthwise case batches a matmul over the channels
so a step scales its own kernel. A single col2im_1d then scatter-adds
the columns back to the signal, with the same shape as before, so the
overlap-add tail, the streaming state and the bias are untouched.

Generation time per frame drops by 80% on CUDA and by 50% on CPU. The
output matches the previous implementation sample for sample, with a
correlation of 0.999994 and identical frame counts.

* flow_temp +  frames_after_eos

* chunking

* mtmd: carry the remaining pocket-tts per-pack settings

The language packs also tune the end-of-speech padding and the padding
of short prompts, next to the temperature already carried in the
mmproj: french_24l asks for 8 tail frames instead of the guessed 3,
english_2026-01 asks for short prompts to be padded with spaces.

Write both in the mmproj as clip.gen.audio.frames_after_eos and
clip.gen.audio.pad_short_text, keyed on the pack in the conversion
script like the temperature. The loader keeps them optional, so a
mmproj without them behaves as before. Map semicolons to commas for
every pack instead, the reference only asks for it on three of them and
it costs nothing elsewhere.

Existing mmproj files must be converted again to carry the two keys.

On a long french text the port now lands within 2% of the reference:
22.96s against 23.44s, with the same peak level and the same amount of
silence.

* clip.gen.audio.model_variant

* clean up code comments

* nit: drop the dead flow_temp hparam, the pack table holds the default

* update docs

* address security problems

* less invasive base.py

* lint

* add mtmd_gen_inp_default

* add docs

* rm gen_flow_temp

---------

Co-authored-by: Pascal <admin@serveurperso.com>

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.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:
@@ -2637,6 +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/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/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);
+}