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>

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;