73 lines
3.1 KiB
Python
73 lines
3.1 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Qwen3-TTS pipeline: Talker (text → RVQ codec) → Code2Wav (codec → audio).
|
|
|
|
Chunked vs end-to-end mode is dispatched from ``deploy.async_chunk``.
|
|
"""
|
|
|
|
from vllm_omni.config.stage_config import (
|
|
PipelineConfig,
|
|
StageExecutionType,
|
|
StagePipelineConfig,
|
|
)
|
|
from vllm_omni.outputs.output_modality import (
|
|
TensorAccumulationStrategy,
|
|
register_key_accumulation_strategy,
|
|
)
|
|
|
|
_PROC = "vllm_omni.model_executor.stage_input_processors.qwen3_tts"
|
|
|
|
# The talker's codec-frame keys are not waveform chunks, so they must not go
|
|
# through the AUDIO modality's waveform-tuned default (CONCAT_LAST) whenever
|
|
# a stage carrying them consolidates its output under that modality (e.g. an
|
|
# AUDIO-typed final stage that forwards them directly instead of routing
|
|
# through code2wav): ``codes.audio`` grows one row per decode step along
|
|
# dim 0, and ``codes.ref`` re-emits the constant reference-context matrix
|
|
# unchanged at every step.
|
|
register_key_accumulation_strategy("codes.audio", TensorAccumulationStrategy.CONCAT_DIM0)
|
|
register_key_accumulation_strategy("codes.ref", TensorAccumulationStrategy.REPLACE)
|
|
|
|
QWEN3_TTS_PIPELINE = PipelineConfig(
|
|
model_type="qwen3_tts",
|
|
default_deploy_config_name="qwen3_tts.yaml",
|
|
# Pipeline-level default; the code2wav stage overrides per-stage below.
|
|
model_arch="Qwen3TTSTalkerForConditionalGeneration",
|
|
stages=(
|
|
StagePipelineConfig(
|
|
stage_id=0,
|
|
model_stage="qwen3_tts",
|
|
execution_type=StageExecutionType.LLM_AR,
|
|
input_sources=(),
|
|
owns_tokenizer=True,
|
|
engine_output_type="latent",
|
|
async_chunk_process_next_stage_input_func=(f"{_PROC}.talker2code2wav_async_chunk"),
|
|
custom_process_next_stage_input_func=f"{_PROC}.talker2code2wav_full_payload",
|
|
sampling_constraints={
|
|
"detokenize": False,
|
|
"stop_token_ids": [2150],
|
|
},
|
|
),
|
|
StagePipelineConfig(
|
|
stage_id=1,
|
|
model_stage="code2wav",
|
|
execution_type=StageExecutionType.LLM_GENERATION,
|
|
input_sources=(0,),
|
|
final_output=True,
|
|
final_output_type="audio",
|
|
engine_output_type="audio",
|
|
model_arch="Qwen3TTSCode2Wav",
|
|
# ``sync_process_input_func`` is the only input-proc override for
|
|
# this stage in sync (non-async-chunk) mode: a length-only
|
|
# ``_token_only`` placeholder. The bulk codec payload itself
|
|
# ships via the worker connector from stage 0's
|
|
# ``talker2code2wav_full_payload`` producer. Under async_chunk
|
|
# mode no pre-stage processing is needed -- chunks deliver
|
|
# directly to the consumer.
|
|
sync_process_input_func=f"{_PROC}.talker2code2wav_token_only",
|
|
sampling_constraints={"detokenize": True},
|
|
extras={"tts_args": {"max_instructions_length": 500}},
|
|
requires_full_payload_input=True,
|
|
),
|
|
),
|
|
)
|