Add ONNX export of ai4bharat/indic-parler-tts (parity verified vs PyTorch).

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+decoder_model.onnx.data filter=lfs diff=lfs merge=lfs -text
+decoder_with_past_model.onnx.data filter=lfs diff=lfs merge=lfs -text
+text_encoder.onnx.data filter=lfs diff=lfs merge=lfs -text

NOTICE.md

@@ -0,0 +1,27 @@
+# Notice
+
+This repository contains an ONNX re-export of `ai4bharat/indic-parler-tts`.
+
+## Provenance
+
+- **Upstream model**: [`ai4bharat/indic-parler-tts`](https://huggingface.co/ai4bharat/indic-parler-tts) by AI4Bharat
+- **Upstream framework**: [`huggingface/parler-tts`](https://github.com/huggingface/parler-tts) by Yoach Lacombe, Vaibhav Srivastav, and Sanchit Gandhi
+- **License**: Apache 2.0 (preserved verbatim from upstream)
+
+## What we did
+
+- Loaded the upstream Safetensors weights via `parler_tts.ParlerTTSForConditionalGeneration.from_pretrained`.
+- Forced `_attn_implementation = "eager"` on every nested config (text_encoder, decoder, audio_encoder) to make `torch.export.export` succeed — SDPA's data-dependent branching is incompatible with dynamo tracing.
+- Wrapped the text encoder, decoder (no-past), and decoder-with-past as three thin `nn.Module`s and exported each via `torch.onnx.export(..., dynamo=True, external_data=True)`.
+- Verified per-graph parity vs PyTorch eager forward (max abs diff `4.72e-6` for text encoder, `1.91e-5` for decoder step 1, `1.68e-4` for decoder step 2).
+
+No model surgery, no quantisation, no weight modification — these ONNX files run the same arithmetic as the upstream Safetensors model in float32.
+
+## What we did NOT include
+
+- The DAC vocoder is bundled inside the upstream model as the `audio_encoder` submodule, but for the browser we consume [`onnx-community/dac_44khz-ONNX`](https://huggingface.co/onnx-community/dac_44khz-ONNX) separately. This avoids duplicate weights in the bundle.
+- The `apply_delay_pattern_mask` autoregressive helper is left out of the ONNX graph (it's stateful in a way that doesn't trace cleanly). Consumers port that ~30-line function to JS — see the upstream `parler_tts/modeling_parler_tts.py` for reference.
+
+## Citation
+
+Please credit AI4Bharat and the Parler-TTS authors as described in the upstream model card.

README.md

@@ -0,0 +1,185 @@
+---
+language:
+- hi
+- ta
+- te
+- bn
+- mr
+- gu
+- kn
+- ml
+- or
+- pa
+- as
+- ur
+- ne
+- sa
+- mai
+- sd
+- kok
+- mni
+- sat
+license: apache-2.0
+library_name: transformers.js
+pipeline_tag: text-to-audio
+tags:
+- onnx
+- text-to-speech
+- tts
+- indic
+- parler-tts
+- onnxruntime-web
+base_model: ai4bharat/indic-parler-tts
+---
+
+# Indic Parler-TTS — ONNX
+
+This is a verbatim ONNX export of [`ai4bharat/indic-parler-tts`](https://huggingface.co/ai4bharat/indic-parler-tts) for in-browser inference via [onnxruntime-web](https://onnxruntime.ai/) or [transformers.js](https://huggingface.co/docs/transformers.js).
+
+The upstream model is the multilingual Indic-language fine-tune of Parler-TTS ([huggingface/parler-tts](https://github.com/huggingface/parler-tts)) by AI4Bharat. This repo re-exports the same weights as ONNX graphs so they can be loaded into a browser without a Python runtime.
+
+## Files
+
+| File | Size | Purpose |
+|---|---|---|
+| `text_encoder.onnx` + `.data` | 1.30 GB | T5 (flan-t5-large) description encoder |
+| `decoder_model.onnx` + `.data` | 1.98 GB | First AR step — codec embed + prompt prefix + cross-attn + 24-layer decoder + fused 9-codebook lm-head |
+| `decoder_with_past_model.onnx` + `.data` | 1.25 GB | Subsequent AR steps with past KV cache |
+| `tokenizer.json`, `tokenizer.model`, `tokenizer_config.json` | 12 MB | LlamaTokenizerFast, vocab 90,714 — used for both description AND prompt |
+| `config.json`, `generation_config.json`, `special_tokens_map.json` | small | Model + generation config |
+
+**Total bundle: ~4.5 GB fp32.** The DAC 44.1 kHz vocoder ([`onnx-community/dac_44khz-ONNX`](https://huggingface.co/onnx-community/dac_44khz-ONNX), referenced separately) is needed to turn the decoder's codec tokens into audio waveform.
+
+## Architecture facts (consumer cheat sheet)
+
+- **Single tokenizer** for description + prompt (the upstream README's mention of a `prompt_tokenizer/` subfolder is a documentation artifact — both inputs use the same `LlamaTokenizerFast`, vocab 90,714).
+- **`prompt_cross_attention=False`** — prompt embeddings are *prefixed* to the decoder's input embeddings, not cross-attended. The decoder's self-attention KV cache region therefore covers `(prompt_len + codec_len)` positions after step 1.
+- **Decoder**: 24 layers × 16 heads × 64 head_dim (1024 hidden), 9 codebooks, vocab 1088 per codebook, max position 4096.
+- **Fused lm_head**: a single `Linear: 1024 → 9 × 1088` projection (not 9 separate heads), reshape + transpose inside the graph.
+- **PAD/BOS/EOS/START**: 1024 / 1025 / 1024 / 1025.
+- **Sample rate**: 44100 Hz from DAC. ~10 ms per codec token end-to-end.
+
+## I/O signatures
+
+### `text_encoder.onnx`
+
+| Input | Shape | Type |
+|---|---|---|
+| `input_ids` | `[B, S]` | int64 |
+| `attention_mask` | `[B, S]` | int64 |
+
+| Output | Shape | Type |
+|---|---|---|
+| `last_hidden_state` | `[B, S, 1024]` | float32 |
+
+### `decoder_model.onnx` (first AR step)
+
+| Input | Shape | Type |
+|---|---|---|
+| `codec_input_ids` | `[B, 9, T]` | int64 |
+| `prompt_input_ids` | `[B, P]` | int64 |
+| `prompt_attention_mask` | `[B, P]` | int64 |
+| `encoder_hidden_states` | `[B, S, 1024]` | float32 |
+| `encoder_attention_mask` | `[B, S]` | int64 |
+
+| Output | Shape | Type |
+|---|---|---|
+| `logits` | `[9*B, P+T, 1088]` | float32 |
+| `present.{i}.{decoder.key,decoder.value,encoder.key,encoder.value}` × 24 layers | `[B, 16, P+T or S, 64]` | float32 |
+
+### `decoder_with_past_model.onnx` (subsequent AR steps)
+
+| Input | Shape | Type |
+|---|---|---|
+| `codec_input_ids` | `[B, 9, 1]` | int64 |
+| `attention_mask` | `[B, full_t]` | int64 — covers prompt + prior codec + new |
+| `encoder_attention_mask` | `[B, S]` | int64 |
+| `cache_position` | `[1]` | int64 — `past_kv_length` scalar |
+| `past_key_values.{i}.{decoder.key,decoder.value,encoder.key,encoder.value}` × 24 layers | as in `present` above | float32 |
+
+| Output | Shape | Type |
+|---|---|---|
+| `logits` | `[9*B, 1, 1088]` | float32 |
+| `present.{i}.{...}` × 24 layers | shifted by 1 in self-attn time dim | float32 |
+
+KV layout per layer: `(decoder_self.key, decoder_self.value, encoder_cross.key, encoder_cross.value)` — IT2-compatible.
+
+## End-to-end inference (sketch)
+
+```js
+// 1. Encode description
+const encOut = await sessText.run({ input_ids, attention_mask });
+const enc_h = encOut.last_hidden_state;
+
+// 2. Build initial codec — [B, 9, 1] all START_ID=1025
+const codec0 = new BigInt64Array(B * 9 * 1).fill(1025n);
+
+// 3. Build delay-pattern mask (port of parler_tts.build_delay_pattern_mask)
+const { initialIds, patternMask } = buildDelayPatternMask({
+  bos: 1025, pad: 1024, maxLen: 256, numCodebooks: 9,
+});
+
+// 4. First step
+let { logits, ...present } = await sessDecNoPast.run({
+  codec_input_ids: initialIds,           // [B, 9, T_init]
+  prompt_input_ids,
+  prompt_attention_mask,
+  encoder_hidden_states: enc_h,
+  encoder_attention_mask: attention_mask,
+});
+
+// 5. Greedy AR loop — apply delay pattern between steps
+for (let step = 1; step < maxSteps; step++) {
+  const next = applyDelayPatternMask(argmaxLogits(logits), patternMask, step);
+  ({ logits, ...present } = await sessDecWithPast.run({
+    codec_input_ids: next,
+    attention_mask: buildFullAttnMask(promptLen + step),
+    encoder_attention_mask,
+    cache_position: BigInt(promptLen + step),
+    ...present,                          // re-feed all 96 KV tensors
+  }));
+}
+
+// 6. Decode codec tokens → audio via DAC
+//    use onnx-community/dac_44khz-ONNX decoder
+```
+
+The `apply_delay_pattern_mask` and `build_delay_pattern_mask` functions need porting to JS — reference the Python source in [`parler_tts/modeling_parler_tts.py`](https://github.com/huggingface/parler-tts/blob/main/parler_tts/modeling_parler_tts.py).
+
+## Validated parity vs PyTorch
+
+- Text encoder: max abs diff `4.72e-6`
+- Decoder step 1 (no-past): max abs diff `1.91e-5`
+- Decoder step 2 (with-past): max abs diff `1.68e-4`
+
+All well below the typical fp32-sufficient `1e-3` tolerance.
+
+## Citation
+
+If you use this in research or production, please cite the original AI4Bharat work:
+
+```bibtex
+@misc{indic-parler-tts,
+  title = {Indic Parler-TTS},
+  author = {AI4Bharat},
+  year = {2024},
+  url = {https://huggingface.co/ai4bharat/indic-parler-tts}
+}
+```
+
+And the upstream Parler-TTS:
+
+```bibtex
+@misc{lacombe-etal-2024-parler-tts,
+  author = {Yoach Lacombe and Vaibhav Srivastav and Sanchit Gandhi},
+  title = {Parler-TTS},
+  year = {2024},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/huggingface/parler-tts}}
+}
+```
+
+## License
+
+Apache 2.0 — same as the upstream `ai4bharat/indic-parler-tts`. See [`NOTICE.md`](NOTICE.md) for attribution.

config.json

@@ -0,0 +1,274 @@
+{
+  "_name_or_path": "/fsx/yoach/tmp/artefacts/training-multilingual-mini-indic-finetuning-on-base/",
+  "architectures": [
+    "ParlerTTSForConditionalGeneration"
+  ],
+  "audio_encoder": {
+    "_name_or_path": "ylacombe/dac_44khz",
+    "add_cross_attention": false,
+    "architectures": [
+      "DacModel"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "codebook_dim": 8,
+    "codebook_loss_weight": 1.0,
+    "codebook_size": 1024,
+    "commitment_loss_weight": 0.25,
+    "cross_attention_hidden_size": null,
+    "decoder_hidden_size": 1536,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "downsampling_ratios": [
+      2,
+      4,
+      8,
+      8
+    ],
+    "early_stopping": false,
+    "encoder_hidden_size": 64,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_size": 1024,
+    "hop_length": 512,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "dac",
+    "n_codebooks": 9,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "quantizer_dropout": 0.0,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sampling_rate": 44100,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "upsampling_ratios": [
+      8,
+      8,
+      4,
+      2
+    ],
+    "use_bfloat16": false
+  },
+  "decoder": {
+    "_name_or_path": "/fsx/yoach/tmp/artefacts/parler-tts-mini-v2-empty/decoder",
+    "activation_dropout": 0.0,
+    "activation_function": "gelu",
+    "add_cross_attention": true,
+    "architectures": [
+      "ParlerTTSForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 1025,
+    "chunk_size_feed_forward": 0,
+    "codebook_weights": null,
+    "cross_attention_hidden_size": null,
+    "cross_attention_implementation_strategy": null,
+    "decoder_start_token_id": null,
+    "delay_strategy": "delay",
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1024,
+    "exponential_decay_length_penalty": null,
+    "ffn_dim": 4096,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layerdrop": 0.0,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 4096,
+    "min_length": 0,
+    "model_type": "parler_tts_decoder",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codebooks": 9,
+    "num_cross_attention_key_value_heads": 16,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 16,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 1024,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rope_embeddings": false,
+    "rope_theta": 10000.0,
+    "scale_embedding": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "use_fused_lm_heads": true,
+    "vocab_size": 1088
+  },
+  "decoder_start_token_id": 1025,
+  "is_encoder_decoder": true,
+  "model_type": "parler_tts",
+  "pad_token_id": 1024,
+  "prompt_cross_attention": false,
+  "text_encoder": {
+    "_name_or_path": "google/flan-t5-large",
+    "add_cross_attention": false,
+    "architectures": [
+      "T5ForConditionalGeneration"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": null,
+    "d_ff": 2816,
+    "d_kv": 64,
+    "d_model": 1024,
+    "decoder_start_token_id": 0,
+    "dense_act_fn": "gelu_new",
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout_rate": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1,
+    "exponential_decay_length_penalty": null,
+    "feed_forward_proj": "gated-gelu",
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 1.0,
+    "is_decoder": false,
+    "is_encoder_decoder": true,
+    "is_gated_act": true,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-06,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "t5",
+    "n_positions": 512,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_decoder_layers": 24,
+    "num_heads": 16,
+    "num_layers": 24,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_past": true,
+    "output_scores": false,
+    "pad_token_id": 0,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "relative_attention_max_distance": 128,
+    "relative_attention_num_buckets": 32,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 32128
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.46.0.dev0",
+  "vocab_size": 90714
+}

decoder_model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e9d50f483015207dba00e0373d1dfb255a5bceb6002e5756c86fe6962700da
+size 3401863

decoder_model.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:029c8a9fd562e0fd4ff0513df3d7cea951277becc759ff497aed0c15d3774c53
+size 2080415744

decoder_with_past_model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9504153915896e9e0ae7d20e8ae7bc558c71606c4f185f5dd13904a5c3350ebf
+size 2246622

decoder_with_past_model.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26b7597ba133486e0c3223b47154b8955f491d6214071c72404039c43e663e2d
+size 1306001408

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1025,
+  "decoder_start_token_id": 1025,
+  "do_sample": true,
+  "eos_token_id": 1024,
+  "max_length": 2610,
+  "pad_token_id": 1024,
+  "transformers_version": "4.46.0.dev0"
+}

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

text_encoder.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:200187fa4148ddbd6b6a3edeb82cf7b4812fece9302a5a1934b3082bb5093bf5
+size 2476571

text_encoder.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fffdcffc587047ce0fd69b6e3577087097decd2b98d7e994d29b9928a206e94
+size 1364983808

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc8fa773221597d09cfadb23a2b1bd717488a0481505469ea56d42cb044de9b5
+size 1795391

tokenizer_config.json

@@ -0,0 +1,43 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<unk>",
+  "padding_side": "left",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}