[Bugfix][Quantization] Detect NVFP4 in ModelOpt mixed-precision checkpoints (#56050)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Stefano Castagnetta
2026-10-02 12:30:22 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent fa212b8dc5
commit f0b4b36890
4 changed files with 125 additions and 1 deletions
+14
View File
@@ -101,6 +101,20 @@ llama3_8b_fp4 = ModelFusionInfo(
),
)
# ModelOpt MIXED_PRECISION: FP8 attention projections, NVFP4 MLPs. It resolves to
# modelopt_mixed, not modelopt_fp4, and its MLP has no manual act+quant fusion, so
# every MLP boundary is left to ActivationQuantFusionPass.
qwen3_8_27b_mixed_fp4 = ModelFusionInfo(
model_name="nvidia/Qwen3.8-27B-NVFP4",
hf_overrides=lambda n_layers: {
"text_config": {
"num_hidden_layers": n_layers,
"layer_types": ["linear_attention"] * (n_layers - 1) + ["full_attention"],
}
},
matches=lambda n_layers: Matches(act_quant_fusion=n_layers),
)
# MoEs cannot do act+quant fusion because those ops are hidden from torch.compile.
# MoEs also only expose 1 rms+quant fusion because the quant for up_proj is hidden.
# TODO(luka): https://github.com/vllm-project/vllm/issues/31985