From aa790870176752359fdc07bc509b09538ca24f81 Mon Sep 17 00:00:00 2001 From: Raymond Huang <103474329+Raymondlol@users.noreply.github.com> Date: Sat, 10 Oct 2026 01:03:44 -0400 Subject: [PATCH] [diffusion] fix: fail MiniMax-H3 quality="high" closed on substituted DiT weights (#34978) Co-authored-by: Claude Opus 5 --- .../configs/pipeline_configs/minimax_h3.py | 4 +++ .../test/unit/test_minimax_h3_admission.py | 27 +++++++++++++++++++ 2 files changed, 31 insertions(+) diff --git a/python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py b/python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py index e4da4079a3..1aeba59e8c 100644 --- a/python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py +++ b/python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py @@ -155,6 +155,7 @@ class MiniMaxH3PipelineConfig(PipelineConfig): "is_dit_layerwise_offload_selected": ( server_args.is_dit_layerwise_offload_selected ), + "lora_path": server_args.lora_path, "model_variant": model_variant, "num_gpus": server_args.num_gpus, "performance_mode": server_args.performance_mode, @@ -165,6 +166,7 @@ class MiniMaxH3PipelineConfig(PipelineConfig): "ring_degree": server_args.ring_degree, "sp_degree": server_args.sp_degree, "tp_size": server_args.tp_size, + "transformer_weights_path": server_args.transformer_weights_path, "ulysses_degree": server_args.ulysses_degree, "use_fsdp_inference": server_args.use_fsdp_inference, } @@ -178,6 +180,7 @@ class MiniMaxH3PipelineConfig(PipelineConfig): "enable_breakable_cuda_graph": False, "enable_torch_compile": False, "is_dit_layerwise_offload_selected": False, + "lora_path": None, "model_variant": "fl2va", "num_gpus": 4, "performance_mode": "speed", @@ -188,6 +191,7 @@ class MiniMaxH3PipelineConfig(PipelineConfig): "ring_degree": 1, "sp_degree": 4, "tp_size": 1, + "transformer_weights_path": None, "ulysses_degree": 4, "use_fsdp_inference": False, } diff --git a/python/sglang/multimodal_gen/test/unit/test_minimax_h3_admission.py b/python/sglang/multimodal_gen/test/unit/test_minimax_h3_admission.py index d8e2dd31ad..2d33b94ac2 100644 --- a/python/sglang/multimodal_gen/test/unit/test_minimax_h3_admission.py +++ b/python/sglang/multimodal_gen/test/unit/test_minimax_h3_admission.py @@ -367,6 +367,7 @@ def _quality_server_args(): enable_breakable_cuda_graph=False, enable_torch_compile=False, is_dit_layerwise_offload_selected=False, + lora_path=None, minimax_h3_adaln_online=False, performance_mode="speed", quantization=None, @@ -558,6 +559,32 @@ def test_quality_admission_fails_closed_outside_validated_request(): stage.forward(batch, server_args) +@pytest.mark.parametrize( + ("field", "value"), + [ + ("lora_path", "larryvrh/MiniMax-H3-Turbo-Lora"), + ("transformer_weights_path", "/weights/dit.safetensors"), + ], +) +def test_quality_admission_rejects_substituted_dit_weights(field, value): + config = MiniMaxH3PipelineConfig() + server_args = _quality_server_args() + + with ( + patch.object(current_platform, "is_cuda", return_value=True), + patch.object(current_platform, "get_device_name", return_value="NVIDIA H200"), + patch.object( + current_platform, + "get_device_capability", + return_value=_HopperCapability(), + ), + ): + config.validate_quality_deployment(server_args) + setattr(server_args, field, value) + with pytest.raises(ValueError, match=field): + config.validate_quality_deployment(server_args) + + def test_validate_server_args_requires_packed_varlen_backend(): config = MiniMaxH3PipelineConfig() server_args = SimpleNamespace(