[NPU] [Diffusion] Fix NPU multimodal-gen CI (#41528)

This commit is contained in:
Артем Савкин
2026-10-01 18:02:37 +03:00
committed by GitHub
parent 41cbe65de0
commit 1bbce57d9e
5 changed files with 27 additions and 11 deletions
@@ -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
@@ -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(
@@ -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,
@@ -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/"
@@ -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(