Files
vllm-omni/vllm_omni/deploy/minicpmo_4_5_duplex_h200.yaml
T

37 lines
1.5 KiB
YAML

# Opt-in single-H200 native full-duplex codec profile (V1).
# Inherits the native session/stage capacities, sampling and chunk sizes.
# Enables all Stage 0 (multimodal duplex prefill) and Stage 2 (code2wav DiT) optimizations.
# Ordinary TF32 for CFM DiT GEMMs (cuBLAS allow_tf32 + Triton tf32 dots).
# HiFT stays IEEE FP32. Validate audio quality for your workload.
base_config: minicpmo_4_5.yaml
connectors:
connector_of_shared_memory:
extra:
token2wav_allow_tf32: true
# Stage 2: Fused DiT transformer body (LayerNorm + QKV + Residual)
cfm_fused_body: true
# Stage 2: Flow condition encoder CUDA graph
cfm_encoder_cuda_graph: true
# Stage 2: Pre-allocated persistent slot pool and row offset merging
cfm_slot_pool: true
cfm_row_offset_merge: true
# Stage 2: Compress DiT attention cache to bf16
code2wav_bfloat16_attention_cache: true
stages:
- stage_id: 0
enforce_eager: false
hf_overrides:
# Stage 0: Incremental sliding window Fbank feature extraction
duplex_incremental_fbank: true
# Stage 0: Fused packed SigLIP encoder layers
vision_fused_layers: true
# Stage 0: Packed SigLIP tower & resampler CUDA graph
vision_cuda_graph: true
# Stage 0: Steady Whisper streaming audio encoder CUDA graph
duplex_audio_encoder_cuda_graph: true
duplex_audio_encoder_cuda_graph_batch_sizes_from_sessions: true
# Stage 0: Pinned memory async H2D copy for streaming audio chunks
duplex_audio_encoder_pinned_h2d: true