From 35608e48d123bb5214ed5b8f3034b2a4e3d36533 Mon Sep 17 00:00:00 2001 From: Mick Date: Thu, 8 Oct 2026 09:24:53 +0800 Subject: [PATCH] [diffusion] refactor: reuse the single-process parallel test fixture (#42964) Co-authored-by: Mick Qian --- .../test/unit/test_flux3_action.py | 30 +++++-------------- .../test/unit/test_mxfp8_online_gemm.py | 29 +++++------------- 2 files changed, 16 insertions(+), 43 deletions(-) diff --git a/python/sglang/multimodal_gen/test/unit/test_flux3_action.py b/python/sglang/multimodal_gen/test/unit/test_flux3_action.py index acb006114d..6d4aba0251 100644 --- a/python/sglang/multimodal_gen/test/unit/test_flux3_action.py +++ b/python/sglang/multimodal_gen/test/unit/test_flux3_action.py @@ -513,20 +513,6 @@ def test_cosmos_unipc_recovers_x0_on_a_straight_flow(): # ---------------------------------------------------------------- DiT -def _init_single_process_parallel() -> None: - from sglang.multimodal_gen.runtime.distributed.parallel_state import ( - maybe_init_distributed_environment_and_model_parallel, - model_parallel_is_initialized, - ) - from sglang.multimodal_gen.test.single_test_file.component_accuracy.utils import ( - ensure_distributed_env_defaults, - ) - - if not model_parallel_is_initialized(): - ensure_distributed_env_defaults() - maybe_init_distributed_environment_and_model_parallel(tp_size=1, sp_size=1) - - def _tiny_arch() -> Flux3ArchConfig: return Flux3ArchConfig( hidden_size=64, @@ -542,12 +528,12 @@ def _tiny_arch() -> Flux3ArchConfig: @pytest.mark.skipif( not torch.cuda.is_available(), reason="needs the CUDA parallel runtime" ) -def test_dit_loads_released_checkpoint_names(): +def test_dit_loads_released_checkpoint_names(request): """Released checkpoints load by name; a renamed module or mapping breaks them.""" from sglang.multimodal_gen.runtime.loader.utils import get_param_names_mapping from sglang.multimodal_gen.runtime.models.dits.flux3 import Flux3Transformer - _init_single_process_parallel() + request.getfixturevalue("single_process_model_parallel") with torch.device("meta"): model = Flux3Transformer(Flux3DiTConfig(arch_config=_tiny_arch())) params = set(model.state_dict()) @@ -594,14 +580,14 @@ def test_dit_loads_released_checkpoint_names(): @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs CUDA attention") -def test_dit_cached_streams_match_full_forward(): +def test_dit_cached_streams_match_full_forward(request): """The pipeline's per-caption / per-request caching must not change predictions.""" from sglang.multimodal_gen.runtime.managers.forward_context import ( set_forward_context, ) from sglang.multimodal_gen.runtime.models.dits.flux3 import Flux3Transformer - _init_single_process_parallel() + request.getfixturevalue("single_process_model_parallel") torch.manual_seed(0) model = Flux3Transformer(Flux3DiTConfig(arch_config=_tiny_arch())).cuda().bfloat16() for p in model.parameters(): @@ -648,7 +634,7 @@ def test_dit_cached_streams_match_full_forward(): @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs CUDA graphs") -def test_cuda_graph_replays_match_eager_denoise_steps(): +def test_cuda_graph_replays_match_eager_denoise_steps(request): """Replays with fresh inputs must match eager steps bit for bit.""" from sglang.multimodal_gen.runtime.managers.forward_context import ( set_forward_context, @@ -656,7 +642,7 @@ def test_cuda_graph_replays_match_eager_denoise_steps(): from sglang.multimodal_gen.runtime.models.dits.flux3 import Flux3Transformer from sglang.multimodal_gen.runtime.vla.cuda_graph import VLATensorGraphRunner - _init_single_process_parallel() + request.getfixturevalue("single_process_model_parallel") torch.manual_seed(0) model = Flux3Transformer(Flux3DiTConfig(arch_config=_tiny_arch())).cuda().bfloat16() for p in model.parameters(): @@ -734,7 +720,7 @@ def test_fp8r_native_format_preserves_values(monkeypatch, fnuz, tuple_output): not torch.cuda.is_available() or torch.cuda.get_device_capability() < (8, 9), reason="needs FP8 scaled_mm", ) -def test_fp8r_checkpoint_loads_fused_rowwise_linears(): +def test_fp8r_checkpoint_loads_fused_rowwise_linears(request): """Native FP8r payloads (E4M3 + per-row scales) must fuse and dequantize consistently.""" from sglang.multimodal_gen.runtime.models.dits.flux3 import ( Flux3Fp8RowwiseLinear, @@ -743,7 +729,7 @@ def test_fp8r_checkpoint_loads_fused_rowwise_linears(): quantize_fp8_rowwise, ) - _init_single_process_parallel() + request.getfixturevalue("single_process_model_parallel") with torch.device("meta"): model = Flux3Transformer(Flux3DiTConfig(arch_config=_tiny_arch())) reference = {} diff --git a/python/sglang/multimodal_gen/test/unit/test_mxfp8_online_gemm.py b/python/sglang/multimodal_gen/test/unit/test_mxfp8_online_gemm.py index 9758d57d24..d267af8f32 100644 --- a/python/sglang/multimodal_gen/test/unit/test_mxfp8_online_gemm.py +++ b/python/sglang/multimodal_gen/test/unit/test_mxfp8_online_gemm.py @@ -11,20 +11,6 @@ import torch pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required") -def _init_parallel() -> None: - from sglang.multimodal_gen.runtime.distributed.parallel_state import ( - maybe_init_distributed_environment_and_model_parallel, - model_parallel_is_initialized, - ) - from sglang.multimodal_gen.test.single_test_file.component_accuracy.utils import ( - ensure_distributed_env_defaults, - ) - - if not model_parallel_is_initialized(): - ensure_distributed_env_defaults() - maybe_init_distributed_environment_and_model_parallel(tp_size=1, sp_size=1) - - def _layer( in_f: int, out_f: int, bias: bool, params_dtype: torch.dtype = torch.bfloat16 ): @@ -41,10 +27,10 @@ def _layer( ).to("cuda") -def test_mxfp8_linear_matches_bf16_and_accepts_prequantized() -> None: +def test_mxfp8_linear_matches_bf16_and_accepts_prequantized(request) -> None: if torch.cuda.get_device_capability()[0] < 10: pytest.skip("cuBLASLt MXFP8 block scaling requires Blackwell or newer") - _init_parallel() + request.getfixturevalue("single_process_model_parallel") from sglang.kernels.ops.diffusion import silu_mul_mxfp8 in_f, out_f, rows = 512, 384, 200 @@ -72,7 +58,6 @@ def test_mxfp8_linear_matches_bf16_and_accepts_prequantized() -> None: def _assert_aligned_layer_falls_back(params_dtype: torch.dtype) -> None: - _init_parallel() layer = _layer(512, 384, bias=False, params_dtype=params_dtype) g = torch.Generator(device="cpu").manual_seed(1) weight = (torch.randn(384, 512, generator=g) * 0.02).to("cuda", params_dtype) @@ -87,23 +72,25 @@ def _assert_aligned_layer_falls_back(params_dtype: torch.dtype) -> None: assert rel < 0.05, rel -def test_pre_blackwell_aligned_layer_falls_back_to_channelwise() -> None: +def test_pre_blackwell_aligned_layer_falls_back_to_channelwise(request) -> None: if torch.cuda.get_device_capability()[0] >= 10: pytest.skip("requires a pre-Blackwell GPU") + request.getfixturevalue("single_process_model_parallel") _assert_aligned_layer_falls_back(torch.bfloat16) -def test_fp16_layer_falls_back_to_channelwise() -> None: +def test_fp16_layer_falls_back_to_channelwise(request) -> None: """The swizzled quantizer takes bf16 only; an fp16 layer must not fail at load.""" + request.getfixturevalue("single_process_model_parallel") _assert_aligned_layer_falls_back(torch.float16) -def test_unaligned_layer_falls_back_to_channelwise() -> None: +def test_unaligned_layer_falls_back_to_channelwise(request) -> None: """The block-scaled GEMM needs K % 32 == 0; such a layer keeps the per-channel fp8 path and still answers a forward. K is 16 rather than a smaller odd size because the fallback's scaled GEMM still wants K % 16 == 0 (ROCm rejects anything else outright).""" - _init_parallel() + request.getfixturevalue("single_process_model_parallel") layer = _layer(16, 128, bias=True) layer.quant_method.process_weights_after_loading(layer) assert not layer.mxfp8 and not layer.quant_method.accepts_mxfp8_input(layer)