From 1bbce57d9e2d7261eadce2f700e063a0647c3268 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=D0=90=D1=80=D1=82=D0=B5=D0=BC=20=D0=A1=D0=B0=D0=B2=D0=BA?= =?UTF-8?q?=D0=B8=D0=BD?= <58187114+OrangeRedeng@users.noreply.github.com> Date: Thu, 1 Oct 2026 18:02:37 +0300 Subject: [PATCH] [NPU] [Diffusion] Fix NPU multimodal-gen CI (#41528) --- .../runtime/disaggregation/scheduler_mixin.py | 15 ++++++--------- .../minimax_h3/stages/text_encoding.py | 1 + .../test/server/ascend/perf_baselines_npu.json | 2 +- .../sglang/multimodal_gen/test/test_utils.py | 2 +- .../test/unit/test_disagg_extra_tensors.py | 18 ++++++++++++++++++ 5 files changed, 27 insertions(+), 11 deletions(-) diff --git a/python/sglang/multimodal_gen/runtime/disaggregation/scheduler_mixin.py b/python/sglang/multimodal_gen/runtime/disaggregation/scheduler_mixin.py index 2001f1fb56..6a0867d393 100644 --- a/python/sglang/multimodal_gen/runtime/disaggregation/scheduler_mixin.py +++ b/python/sglang/multimodal_gen/runtime/disaggregation/scheduler_mixin.py @@ -17,6 +17,7 @@ import pickle import queue import threading import time +from enum import Enum from typing import TYPE_CHECKING, Any import numpy as np @@ -177,20 +178,16 @@ def _is_tensor_like(value) -> bool: def _to_json_serializable(value): + if isinstance(value, Enum): + return value.name if isinstance(value, (torch.Tensor, np.ndarray)): return value.tolist() if isinstance(value, np.generic): return value.item() + if isinstance(value, dict): + return {key: _to_json_serializable(item) for key, item in value.items()} if isinstance(value, (list, tuple)): - converted = [] - for item in value: - if isinstance(item, (torch.Tensor, np.ndarray)): - converted.append(item.tolist()) - elif isinstance(item, np.generic): - converted.append(item.item()) - else: - converted.append(item) - return converted + return [_to_json_serializable(item) for item in value] return value diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/text_encoding.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/text_encoding.py index bbf2e03c63..1ccb2555c4 100644 --- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/text_encoding.py +++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/text_encoding.py @@ -286,6 +286,7 @@ class MiniMaxH3TextEncodingStage(TextEncodingStage): raise ValueError( "MiniMaxH3TextEncodingStage direct encode requires a tokenizer component" ) + self._begin_text_encoder_use(0) with set_forward_context(current_timestep=0, attn_metadata=None): if plan.task == "ref2va": embeddings = self._encode_ref2va( diff --git a/python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json b/python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json index 41dda2fd48..e7f2f28f6c 100644 --- a/python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json +++ b/python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json @@ -708,7 +708,7 @@ "stages_ms": { "InputValidationStage": 0.1, "MiniMaxH3PartitionAdmissionStage": 0.06, - "MiniMaxH3TextEncodingStage": 134.59, + "MiniMaxH3TextEncodingStage": 421.11, "MiniMaxH3VisualEncodingStage": 0.06, "MiniMaxH3AudioEncodingStage": 0.11, "MiniMaxH3LatentPreparationStage": 41.31, diff --git a/python/sglang/multimodal_gen/test/test_utils.py b/python/sglang/multimodal_gen/test/test_utils.py index 16620c2520..f7196de4d5 100644 --- a/python/sglang/multimodal_gen/test/test_utils.py +++ b/python/sglang/multimodal_gen/test/test_utils.py @@ -45,7 +45,7 @@ SGL_TEST_FILES_CI_DATA_REVISION = "dbb70135da54b1f2b172886b28e813d2bcdd16c1" # The NPU pin is kept as a separate branch so ascend GT can be bumped independently # when it's regenerated on its own cadence. if current_platform.is_npu(): - SGL_TEST_FILES_CI_DATA_REVISION = "cbeaa640573a4bb08dc1c41af92221bfae51b7f6" + SGL_TEST_FILES_CI_DATA_REVISION = "609c4cef31864d75dbd11fe3c26b8ae09469204d" SGL_TEST_FILES_CONSISTENCY_GT_ROOT = ( "https://raw.githubusercontent.com/" diff --git a/python/sglang/multimodal_gen/test/unit/test_disagg_extra_tensors.py b/python/sglang/multimodal_gen/test/unit/test_disagg_extra_tensors.py index ac845acd18..83e064e50f 100644 --- a/python/sglang/multimodal_gen/test/unit/test_disagg_extra_tensors.py +++ b/python/sglang/multimodal_gen/test/unit/test_disagg_extra_tensors.py @@ -3,6 +3,8 @@ import pytest import torch +from sglang.multimodal_gen.configs.sample.sampling_params import SamplingParams +from sglang.multimodal_gen.configs.task_type import ModelTaskType from sglang.multimodal_gen.runtime.disaggregation.roles import ( RoleType, filter_modules_for_role, @@ -20,6 +22,22 @@ from sglang.multimodal_gen.runtime.pipelines.qwen_image21 import QwenImage21Pipe from sglang.multimodal_gen.runtime.pipelines_core import Req +def test_task_type_survives_disagg_transfer(): + """A GLM request must cross the JSON transport after AR completes.""" + req = Req( + request_id="glm-transfer", + prompt="test", + sampling_params=SamplingParams(task_type=ModelTaskType.TI2I), + ) + + tensors, scalars = extract_transfer_fields(req) + metadata, buffers = pack_tensors(tensors, scalars) + received, scalars = unpack_tensors([metadata, *[w._view for w in buffers]]) + rebuilt = SchedulerDisaggMixin._build_disagg_req(None, scalars, received) + + assert ModelTaskType.parse(rebuilt.task_type) == ModelTaskType.TI2I + + def test_qwen21_disagg_encoder_loads_condition_vae(): pipeline = object.__new__(QwenImage21Pipeline) modules = filter_modules_for_role(