speechlm2
Use for NeMo SpeechLM2/SALM/SALMAutomodel, duplex STT/S2S/EAR-TTS, Nemotron VoiceChat, HuggingFace export/conversion, vLLM plugin, Automodel backends, and voice-agent troubleshooting.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- speechlm2
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use for NeMo SpeechLM2/SALM/SALMAutomodel, duplex STT/S2S/EAR-TTS, Nemotron VoiceChat, HuggingFace export/conversion, vLLM plugin, Automodel backends, and voice-agent troubleshooting.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/checkpoints-and-export.md10.1 KB
- 参考文档references/configuration.md11.6 KB
- 参考文档references/troubleshooting.md10.8 KB
- 参考文档references/workflows.md12.4 KB
- 脚本scripts/check_speechlm2_config.py12.7 KB
03 SKILL.md 正文
SpeechLM2
Use this sub-skill for NeMo SpeechLM2 speech-LLM work: SALM/SALMAutomodel, duplex STT and speech-to-speech models, Duplex EAR-TTS inside SpeechLM2 pipelines, Nemotron VoiceChat, HuggingFace/safetensors export, the NeMo SpeechLM vLLM plugin, Automodel optional compiled backends, and local voice-agent configuration.
This guidance is self-contained. It distills repository evidence from docs/source/speechlm2/intro.rst, docs/source/speechlm2/models.rst, docs/source/speechlm2/datasets.rst, docs/source/speechlm2/configs.rst, docs/source/speechlm2/training_and_scaling.rst, docs/source/checkpoints/intro.rst, examples/speechlm2/*.py, examples/speechlm2/conf/*.yaml, examples/voice_agent/**, nemo/collections/speechlm2/**, nemo/agents/voice_agent/**, and tests/collections/speechlm2/**.
Route Here When
- The task mentions
nemo.collections.speechlm2,SALM,SALMAutomodel,SALMWithAsrDecoder,DuplexS2SModel,DuplexS2SSpeechDecoderModel,DuplexSTTModel,DuplexEARTTS, orNemotronVoiceChat. - The user needs speech+LLM generation, SpeechLM2 training configs, Lhotse conversation datasets, duplex speech-to-speech/STT evaluation, or SpeechLM2 inference with audio placeholder tokens.
- The task involves SpeechLM2 HuggingFace export,
model.safetensors, distributed checkpoint consolidation,NemotronVoiceChat.from_pretrained, or vLLM serving withnemo_speechlm. - The task involves NeMo voice-agent server/client configuration, tool calling, vLLM/HF LLM backend selection, streaming ASR/TTS service composition, or backchannel/turn-taking settings.
Route Elsewhere
- Classic ASR transcription, ASR fine-tuning, CTC/RNNT decoding, Parakeet/Canary model recipes, or ASR manifests belong in
../asr/SKILL.mdunless SpeechLM2 is explicitly involved. - TTS-only FastPitch, HiFiGAN, Magpie, Kokoro, or speech synthesis outside a SpeechLM2/Nemotron VoiceChat/voice-agent pipeline belongs in
../tts/SKILL.md. - Repo development, test authoring, packaging, lint, CI, and source-tree maintenance belong in the repo-development sub-skill when available.
First Steps
- Identify the workflow: SALM generation/eval, SALM/SALMAutomodel training, duplex STT/S2S, Duplex EAR-TTS inference, Nemotron VoiceChat, HF export, vLLM serving, or voice-agent deployment.
- Read
references/workflows.mdfor the concrete API/command-shape workflow and safe adaptation notes. - Read
references/configuration.mdbefore editing YAML, Hydra overrides, dataset paths, conversation roles, parallelism, or voice-agent config. - Read
references/checkpoints-and-export.mdbefore loading, resuming, converting, exporting, or serving checkpoints. - Read
references/troubleshooting.mdwhen imports, CUDA, compiled backends, data schemas, Hydra overrides, vLLM, or voice-agent services fail. - Run
scripts/check_speechlm2_config.py --helpand then use it on user-provided SpeechLM2 YAML/JSON configs before long-running jobs.
Core Facts
- SpeechLM2 is active development code for augmenting pretrained LLMs with speech understanding and speech generation; expect APIs/configs to change more often than classic ASR/TTS.
SALMuses HuggingFace Transformers for the LLM backbone;SALMAutomodeluses NeMo Automodel and is the right path for FSDP2, TP, PP, CP, EP/MoE, HSDP, native Automodel LoRA, and distributed inference.- Distributed SALM generation/evaluation requires
SALMAutomodel; plainSALMshould stay single-process/single-device for inference. NemotronVoiceChatis inference-only. Train or fine-tune its STT and EAR-TTS components separately, then export a joint VoiceChat checkpoint.- SpeechLM2 model checkpoints are HuggingFace-style directories with
config.jsonandmodel.safetensors;.nemois used in this collection mainly as an input format for pretrained ASR/perception components. - Training and most inference require a compatible GPU/CUDA/PyTorch stack. Optional Automodel compiled backends such as TransformerEngine, DeepEP, grouped GEMM, flash-attn, vLLM, Pipecat, browser/client tooling, and Node/npm must be installed only when that workflow actually needs them.
Bundled Assets
references/workflows.md: SALM/SALMAutomodel generation and training, duplex STT/S2S/EAR-TTS, Nemotron VoiceChat, vLLM, and voice-agent workflows.references/configuration.md: SpeechLM2 YAML structure, data schemas, Hydra overrides, parallelism, and voice-agent configuration patterns.references/checkpoints-and-export.md: checkpoint formats,from_pretrained,save_pretrained, HF export, vLLM readiness, and Nemotron VoiceChat conversion.references/troubleshooting.md: actionable failure diagnosis for install/import, CUDA/backends, data/config, API/CLI, export, vLLM, and voice-agent issues.scripts/check_speechlm2_config.py: deterministic local checker for SpeechLM2 YAML/JSON-like configs; it performs text-level validation without importing NeMo or downloading models.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).