From d99adcebdc9402c26346230ef2480d1cb72e654e Mon Sep 17 00:00:00 2001 From: Jason Li Date: Tue, 7 Jul 2026 18:19:42 -0400 Subject: [PATCH] [BugFix] Fix ModelOpt quantization inference for fused siblings (#47445) Signed-off-by: jasonlizhengjian --- tests/quantization/test_modelopt.py | 32 ++++++++++++++++++++++++++ vllm/models/minimax_m3/nvidia/model.py | 10 ++++++++ 2 files changed, 42 insertions(+) diff --git a/tests/quantization/test_modelopt.py b/tests/quantization/test_modelopt.py index f10b3c9a55..a08e14c53c 100644 --- a/tests/quantization/test_modelopt.py +++ b/tests/quantization/test_modelopt.py @@ -133,6 +133,38 @@ def test_modelopt_mixed_precision_quantizes_parallel_lm_head(): assert isinstance(method, ModelOptNvFp4LinearMethod) +def test_modelopt_mixed_precision_resolves_declared_packed_projection(): + config = _mixed_precision_config( + { + "model.layers.0.self_attn.q_proj": {"quant_algo": "MXFP8"}, + "model.layers.0.self_attn.k_proj": {"quant_algo": "MXFP8"}, + "model.layers.0.self_attn.v_proj": {"quant_algo": "MXFP8"}, + } + ) + config.packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]} + + assert config._resolve_quant_algo("model.layers.0.self_attn.qkv_proj") == "MXFP8" + + +def test_modelopt_mixed_precision_does_not_quantize_unlisted_fused_sibling(): + config = _mixed_precision_config( + { + "model.layers.0.linear_attn.in_proj_qkv": {"quant_algo": "FP8"}, + "model.layers.0.linear_attn.in_proj_z": {"quant_algo": "FP8"}, + "model.layers.0.linear_attn.out_proj": {"quant_algo": "FP8"}, + } + ) + config.packed_modules_mapping = { + "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"], + "in_proj_ba": ["in_proj_b", "in_proj_a"], + } + + assert ( + config._resolve_quant_algo("model.layers.0.linear_attn.in_proj_qkvz") == "FP8" + ) + assert config._resolve_quant_algo("model.layers.0.linear_attn.in_proj_ba") is None + + def test_modelopt_mixed_precision_infers_fused_gate_up_projection(): from vllm.model_executor.layers.linear import LinearBase diff --git a/vllm/models/minimax_m3/nvidia/model.py b/vllm/models/minimax_m3/nvidia/model.py index 0b79eeb2af..2c59fb0af3 100644 --- a/vllm/models/minimax_m3/nvidia/model.py +++ b/vllm/models/minimax_m3/nvidia/model.py @@ -957,6 +957,11 @@ class MiniMaxM3Model(nn.Module, EagleModelMixin): class MiniMaxM3SparseForCausalLM(nn.Module, SupportsPP, SupportsEagle3): """MiniMax M3 (sparse/dense backbone) for causal language modeling.""" + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], + } + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_text_config @@ -1024,6 +1029,11 @@ class MiniMaxM3SparseForConditionalGeneration( # ranks (see ``_process_image_input`` / ``_process_video_input``). supports_encoder_tp_data = True + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], + } + hf_to_vllm_mapper = WeightsMapper( orig_to_new_prefix={ "multi_modal_projector.": "vision_tower.multi_modal_projector.",