Ming-3.0-ASR-tiny native MTP draft
A one-layer native MTP draft fine-tuned to match the frozen inclusionAI/Ming-3.0-asr-tiny verifier, pinned to revision ac95e7b1678c08ea6f4a2581a051e432ed63e9b7. The repository stores MTP-only weights. Serving acceptance length, throughput, English WER and Chinese CER have been measured in SGLang-Omni; see the online serving evaluation for results and limitations. That integration loads this prediction block automatically when NEXTN is enabled, using shared weights from the original verifier.
Configuration
| Setting | Value |
|---|---|
| Native MTP layers | 1 |
| Native checkpoint layer index | 24 |
| Hidden size / vocabulary size | 1,536 / 157,184 |
| Trainable parameters | 316,604,672 |
| Training rollout | 3 recurrent draft steps through the same layer |
| Checkpoint weights | BF16 |
The weight file contains only the 403 native model.layers.24.* tensors. Shared verifier embeddings and LM-head weights are obtained separately from the pinned base model. Use the verifier's tokenizer and audio processor. The configuration retains the target depth of 24 so the native loader addresses layer 24; num_nextn_predict_layers=1 specifies the draft depth. The full vocabulary is retained, so no draft-vocabulary mapping files are required.
Training
The run completed 6 full epochs on two H200 GPUs in approximately 5.60 hours, including validation and checkpointing. The published checkpoint is epoch 3, selected by the lowest held-out validation loss. Training had a minimum of 6 epochs, a maximum of 10, and early stopping after two completed validation epochs without a new minimum once the minimum training length was reached.
The input cohort reuses the selection behind Ming-3.0-asr-tiny-dspark: LibriSpeech train-clean-360, People's Speech clean, VoxPopuli English, official AISHELL-1 train/dev, and FLEURS English/Chinese/Japanese. The current capture retained 356,954 training clips (999.51 audio hours) and 2,575 validation clips after excluding 52 additional non-stopping teacher generations. Emilia-YODAS and Common Voice 17 remain excluded.
| Language | Training clips | Audio hours | Validation clips |
|---|---|---|---|
| en | 231,515 | 832.32 | 1,416 |
| zh | 123,340 | 160.54 | 921 |
| ja | 2,099 | 6.64 | 238 |
Labels are the frozen verifier's greedy transcripts, including the language prefix and EOS. Audio-conditioned sequences are replayed through vLLM to capture actual input embeddings at boundary 0 and final pre-norm states at boundary 24. Prompt, audio, padding, and cross-document targets are excluded from the loss. The native Ming MTP block is initialized from the verifier checkpoint; the verifier, shared embeddings, LM head, and verifier final norm remain frozen.
Fine-tuning uses Speculators with a native Ming adapter, weighted three-step cross entropy (step weights proportional to 1, 0.6, 0.36), fused AdamW at 3e-5, weight decay 0.01, cosine decay with 3% warmup, and 16,384 packed input tokens per GPU. Trainable parameters use FP32 master precision with BF16 activations; framework checkpoints are stored in BF16. Packing combines independent short clips and does not imply a 16K-token recording.
Training and validation use separate official partitions. Exact canonical 16 kHz PCM16 overlap checks found no train/validation or train/SeedTTS audio duplicates. These checks do not establish speaker disjointness or detect near-duplicates. SeedTTS was held out until checkpoint selection finished.
Offline validation
| Checkpoint | Weighted CE loss | Correct prefix ≥1 | Correct prefix ≥2 | Correct prefix ≥3 |
|---|---|---|---|---|
| Initial native MTP | 5.531582 | 37.30% | 3.39% | 0.17% |
| Selected epoch 3 | 0.269103 | 93.65% | 84.88% | 77.38% |
These are cached, teacher-conditioned validation measurements. Serving acceptance is measured separately below.
| Completed epoch | Validation loss |
|---|---|
| 1 | 0.318932 |
| 2 | 0.281512 |
| 3 (selected) | 0.269103 |
| 4 | 0.280994 |
| 5 | 0.290502 |
| 6 | 0.284293 |
Online serving evaluation
English ASR in normal (non-deterministic) mode on one H200 (BF16, TP1), evaluated on SeedTTS:
| Concurrency | Throughput gain | AL (tokens/verify) |
|---|---|---|
| c1 | +17.58% | 3.3238 |
| c8 | +14.28% | 3.3188 |
| c16 | +10.28% | 3.3166 |
See SGLang-Omni PR #3887 for full results, including accuracy, deterministic mode and Chinese ASR.
Provenance and licenses
Training used Speculators at 988ed16b1f91315942049da3093d88dba70c45d9 and vLLM feature capture at dd4e40142a3280ef404493e5117bde0c60c6e389, with the audio-preserving capture support described in vLLM #66050 and #54174. Online serving evaluation uses the SGLang-Omni integration linked above.
publication-manifest.json records parent revision, checkpoint selection, aggregate evaluation results, and SHA-256 hashes. The repository contains inference weights, required configuration/loading files, documentation, and licenses. Training scripts, private adapters, optimizer states, raw logs, audio, and feature caches are maintained separately.
The Ming base model's MIT notice is retained, together with Apache-2.0 and the original source-file notices. Dataset licenses remain those of their respective sources.