[Bugfix] Align Hub kernel dependency guidance

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

docs/user_guide/diffusion/attention_backends/huggingface_hub.md

@@ -12,7 +12,7 @@ implementations to reduce numerical drift and sampling divergence.
 ## Installation
 
 ```bash
-pip install kernels==0.14.1
+pip install kernels==0.15.2
 ```
 
 If `kernels` is unavailable, an explicit `FLASH_ATTN_HUB` or

examples/offline_inference/qwen2_5_omni/end2end.py

@@ -21,7 +21,6 @@ from vllm.multimodal.media.audio import load_audio
 from vllm.sampling_params import SamplingParams
 
 from vllm_omni.entrypoints.omni import Omni
-from vllm_omni.utils.profiler import add_profiler_config_arg
 from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 
 SEED = 42
@@ -326,7 +325,6 @@ def main(args):
     else:
         query_result = query_func()
     args.quantization_config = quantization_config
-    # ``vars(args)`` below forwards profiler_config to Omni.
     omni_kwargs = vars(args).copy()
     # Override CLI --model with the derived model_name.
     omni_kwargs["model"] = model_name
@@ -380,11 +378,11 @@ def main(args):
         for i, prompt in enumerate(prompts):
             prompt["modalities"] = output_modalities
 
-    profiler_enabled = args.profiler_config is not None
+    profiler_enabled = bool(os.getenv("VLLM_TORCH_PROFILER_DIR"))
     if profiler_enabled and hasattr(omni, "start_profile"):
         omni.start_profile(stages=[0])
     elif profiler_enabled:
-        print("[Warn] --profiler-config was provided, but current engine does not support profiler controls.")
+        print("[Warn] VLLM_TORCH_PROFILER_DIR is set, but current engine does not support profiler controls.")
     omni_generator = omni.generate(prompts, sampling_params_list, py_generator=args.py_generator)
 
     # Determine output directory: prefer --output-dir; fallback to --output-wav
@@ -460,7 +458,6 @@ def parse_args():
         default=False,
         help="Enable writing detailed statistics (default: disabled)",
     )
-    add_profiler_config_arg(parser)
     parser.add_argument(
         "--stage-init-timeout",
         type=int,

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

@@ -14,7 +14,6 @@ from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.models.cosyvoice3.tokenizer import get_qwen_tokenizer
 from vllm_omni.model_executor.models.cosyvoice3.utils import extract_text_token
 from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
-from vllm_omni.utils.profiler import add_profiler_config_arg
 
 # Upstream zero-shot reference clip
 ZERO_SHOT_PROMPT_URL = "https://raw.githubusercontent.com/FunAudioLLM/CosyVoice/main/asset/zero_shot_prompt.wav"
@@ -46,7 +45,6 @@ def run_e2e():
         help="Override the deploy config path. If unset, auto-loads "
         "vllm_omni/deploy/cosyvoice3.yaml based on the HF model_type.",
     )
-    add_profiler_config_arg(parser)
     parser.add_argument("--text", type=str, default="Hello, this is a test of the CosyVoice system capability.")
     parser.add_argument(
         "--prompt-text",
@@ -81,7 +79,6 @@ def run_e2e():
         deploy_config=args.deploy_config,
         tokenizer=args.tokenizer,
         log_stats=True,
-        profiler_config=args.profiler_config,
     )
 
     sampling_cfg = {"top_p": 0.8, "top_k": 25, "eos_token_id": 6561 + 1}
@@ -152,8 +149,8 @@ def run_e2e():
 
     sampling_params_list = [gpt_sampling, s2mel_sampling]
 
-    profiler_enabled = args.profiler_config is not None
+    # Start profiling (requires VLLM_TORCH_PROFILER_DIR env var)
-    if profiler_enabled:
+    if os.environ.get("VLLM_TORCH_PROFILER_DIR"):
         print("Starting profiler...")
         omni.start_profile()
 
@@ -161,7 +158,7 @@ def run_e2e():
     outputs = list(omni.generate(prompts, sampling_params_list=sampling_params_list[:2]))
 
     # Stop profiling and get results
-    if profiler_enabled:
+    if os.environ.get("VLLM_TORCH_PROFILER_DIR"):
         print("Stopping profiler...")
         profile_results = omni.stop_profile()
         print(f"Profile traces saved to: {profile_results}")

requirements/common.txt

@@ -13,8 +13,9 @@ transformers >= 5.10.1, < 5.15
 av>=14.0.0
 omegaconf>=2.3.0
 diffusers==0.40.0
-# Hub kernels, including LTX-2.5 NATTEN
+# Hub kernels, including LTX-2.5 NATTEN. Keep this pin in sync with the
-kernels==0.14.1
+# supported transformers range above.
+kernels==0.15.2
 safetensors>=0.8.0
 accelerate==1.12.0
 soundfile>=0.13.1

tests/diffusion/models/ltx2/test_ltx25_pipeline.py

@@ -386,7 +386,7 @@ def test_ltx25_natten_failure_has_actionable_omni_remedy(monkeypatch):
         MissingNattenProcessor,
     )
 
-    with pytest.raises(RuntimeError, match=r"kernels==0\.14\.1.*supported GPU.*allow Hub access"):
+    with pytest.raises(RuntimeError, match=r"kernels==0\.15\.2.*supported GPU.*allow Hub access"):
         ltx2_components._create_ltx25_natten_processor()
 
 

tests/docs/test_offline_profiler_config.py

@@ -1,23 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-
-from pathlib import Path
-
-import pytest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-_REPO_ROOT = Path(__file__).resolve().parents[2]
-_EXAMPLES = (
-    "examples/offline_inference/text_to_speech/cosyvoice3/end2end.py",
-    "examples/offline_inference/qwen2_5_omni/end2end.py",
-)
-
-
-@pytest.mark.parametrize("relative_path", _EXAMPLES)
-def test_offline_examples_use_typed_profiler_config(relative_path: str):
-    source = (_REPO_ROOT / relative_path).read_text(encoding="utf-8")
-
-    assert "add_profiler_config_arg(parser)" in source
-    assert "profiler_enabled = args.profiler_config is not None" in source
-    assert "VLLM_TORCH_PROFILER_DIR" not in source

vllm_omni/diffusion/models/ltx2/ltx2_components.py

@@ -298,7 +298,7 @@ def _create_ltx25_natten_processor() -> LTX2VideoVaeNeighborhoodNattenProcessor:
     except (ImportError, OSError, RuntimeError, ValueError) as exc:
         raise RuntimeError(
             "LTX-2.5 DiffVAE requires the shi-labs/natten Hub kernel. "
-            "Install kernels==0.14.1, use a supported GPU, leave "
+            "Install kernels==0.15.2, use a supported GPU, leave "
             "DIFFUSERS_DISABLE_REMOTE_CODE unset, and allow Hub access during kernel initialization."
         ) from exc
 

vllm_omni/utils/profiler.py

@@ -1,37 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-
-from __future__ import annotations
-
-import argparse
-import json
-from typing import Any
-
-
-def parse_json_object(value: str, flag_name: str = "argument") -> dict[str, Any]:
-    """Parse a CLI value as a JSON object, attributing errors to ``flag_name``."""
-    try:
-        config = json.loads(value)
-    except json.JSONDecodeError as e:
-        raise argparse.ArgumentTypeError(f"{flag_name} must be valid JSON: {e}") from e
-    if not isinstance(config, dict):
-        raise argparse.ArgumentTypeError(f"{flag_name} must be a JSON object")
-    return config
-
-
-def parse_profiler_config(value: str) -> dict[str, Any]:
-    """Parse the JSON object passed to ``--profiler-config``."""
-    return parse_json_object(value, flag_name="--profiler-config")
-
-
-def add_profiler_config_arg(parser: argparse.ArgumentParser) -> None:
-    """Add the shared JSON profiler configuration option to ``parser``."""
-    parser.add_argument(
-        "--profiler-config",
-        type=parse_profiler_config,
-        default=None,
-        help=(
-            "JSON profiler config for torch/cuda profiling, e.g. "
-            '\'{"profiler":"torch","torch_profiler_dir":"./perf"}\'.'
-        ),
-    )