anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity
Files
c4e61133fbb5587e0cf8972ce32ffac4e37fa739
vllm-omni/vllm_omni/config/environment_variable_inventory.py
T
tzhouamandClaude Opus 5 c4e61133fb [Edge][CPU] Fused norms off by default, and register the switches this branch added 
Two corrections and the paperwork they imply.

The fused-RMSNorm registration defaulted to *on*, which contradicts the
measurement it was committed with: those kernels are 3.6x and 4.5x faster in
isolation but measured 72.7 tok/s against 74.0 end to end, because inductor
already fuses the native norm into the residual add and an opaque call breaks
that fusion. The implementations stay registered, so the provider remains
selectable through `kernel_config.ir_op_priority` on a machine where the trade
lands differently; only the default ranking is left alone now, and the test
asserts both states rather than just the opt-in one.

The four switches this branch added (VLLM_OMNI_AOT_CACHE_SHIM,
_CPU_FUSED_NORMS, _CPU_RELEASE_MEMORY, _SPARK_GATE_REDUCTION) were not in the
reviewed environment-variable inventory, so
`test_statically_resolvable_environment_accesses_are_classified` failed. They
are classified now, the reviewed count moves 28 -> 32 with a note saying why,
and each has a row in docs/configuration/environment_variables.md carrying the
measurement that justifies its default.

`VLLM_OMNI_CPU_FAST_SAMPLER` and `VLLM_OMNI_STEP_STATS_SYNC` remain
unclassified and that test still fails on them. They arrived in 6ce8de1b, in
files this branch does not touch, so they are left to whoever owns that work
rather than documented from a guess.

Under `-m "core_model and cpu"`: 2593 passed before these fixes with 9
failures, of which two were mine. The rest are CUDA-only tests in a CPU venv
("Torch not compiled with CUDA enabled") plus test_qnn_export.py, which
another session is actively editing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-12 14:55:24 +08:00

311 lines
9.7 KiB
Python
Raw Blame History

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
"""Reviewed inventory of environment variables used by vLLM-Omni.
This module classifies environment-variable names; it does not read them or
make every classified name part of the public configuration contract. In
particular, model-specific variables remain transitional until their recorded
disposition is implemented.
"""
from __future__ import annotations
from collections.abc import Mapping
from dataclasses import dataclass
from enum import Enum
from types import MappingProxyType
class EnvironmentVariableCategory(str, Enum):
"""Ownership and documentation boundary for an environment variable."""
PUBLIC_OMNI = "public_omni"
INHERITED_VLLM = "inherited_vllm"
PLATFORM_EXTERNAL = "platform_external"
MODEL_SPECIFIC = "model_specific"
BENCHMARK_TRANSITIONAL = "benchmark_transitional"
INTERNAL = "internal"
class ModelEnvironmentVariableDisposition(str, Enum):
"""Required destination for a model-specific environment variable."""
PROMOTE = "promote"
REQUEST_SCOPE = "request_scope"
EXTERNAL = "external"
INTERNALIZE = "internalize"
DEPRECATE_REMOVE = "deprecate_remove"
@dataclass(frozen=True)
class EnvironmentVariableClassification:
"""Classification metadata for one statically identifiable name."""
name: str
category: EnvironmentVariableCategory
model_disposition: ModelEnvironmentVariableDisposition | None = None
redact_value: bool = False
@property
def is_public_omni(self) -> bool:
return self.category is EnvironmentVariableCategory.PUBLIC_OMNI
# New entries must use the VLLM_OMNI_ prefix. The regression test explicitly
# grandfathers the older public names below.
_PUBLIC_OMNI = (
"DIFFUSION_ATTENTION_BACKEND",
"DIFFUSION_CACHE_ADAPTER",
"DIFFUSION_CACHE_BACKEND",
"OMNI_DIFFUSION_PROMPT_EMBED_CACHE",
"OMNI_DIFFUSION_PROMPT_EMBED_CACHE_SIZE",
"OMNI_DIFFUSION_SESSION_STATE_MANAGER",
"OMNI_DIFFUSION_SESSION_STATE_MANAGER_MAX_SESSIONS",
"SPEAKER_MAX_UPLOADED",
"SPEAKER_SAMPLES_DIR",
"VLLM_OMNI_ASYNC_OUTPUT_TIMEOUT",
# edge branch (vllm_omni/edge, init timeline, CPU platform)
"VLLM_OMNI_AOT_CACHE_SHIM",
"VLLM_OMNI_CPU_FUSED_NORMS",
"VLLM_OMNI_CPU_INDUCTOR",
"VLLM_OMNI_CPU_RELEASE_MEMORY",
"VLLM_OMNI_HW_PROFILE",
"VLLM_OMNI_INIT_TIMELINE",
"VLLM_OMNI_STEP_STATS_DIR",
"VLLM_OMNI_EVENT_DRIVEN_ORCH",
"VLLM_OMNI_INPUT_WAIT_TIMEOUT_S",
"VLLM_OMNI_ORCH_MONITOR_PATH",
"VLLM_OMNI_SERVER_STORAGE__FILE_CONCURRENCY",
"VLLM_OMNI_SERVER_STORAGE__FILE_TTL",
"VLLM_OMNI_SERVER_STORAGE__PATH",
"VLLM_OMNI_SERVER_STORAGE__TTL_SWEEP_INTERVAL",
"VLLM_OMNI_SKIP_NVFP4_NAN_CLAMP",
"VLLM_OMNI_SPARK_GATE_REDUCTION",
"VLLM_OMNI_STORAGE_MAX_CONCURRENCY",
"VLLM_OMNI_STORAGE_PATH",
"VLLM_OMNI_USE_QUACK_FP8",
"VLLM_OMNI_VIDEO_SYNC_TIMEOUT",
"VLLM_VIDEO_ASYNC_CHUNK",
"VLLM_VIDEO_AUDIO_DELTA_MODE",
)
_INHERITED_VLLM = (
"CUDA_HOME",
"CUDA_VISIBLE_DEVICES",
"LOCAL_RANK",
"NO_COLOR",
"VLLM_ALLOW_LONG_MAX_MODEL_LEN",
"VLLM_BATCH_INVARIANT",
"VLLM_CACHE_ROOT",
"VLLM_DISABLE_LOG_LOGO",
"VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE",
"VLLM_HTTP_TIMEOUT_KEEP_ALIVE",
"VLLM_LOGGING_COLOR",
"VLLM_MOONCAKE_BOOTSTRAP_PORT",
"VLLM_PLUGINS",
"VLLM_ROCM_USE_AITER",
"VLLM_ROCM_USE_AITER_RMSNORM",
"VLLM_TUNED_CONFIG_FOLDER",
"VLLM_USE_MODELSCOPE",
"VLLM_USE_OINK_OPS",
"VLLM_WORKER_MULTIPROC_METHOD",
"VLLM_XPU_USE_SAMPLER_KERNEL",
)
_PLATFORM_EXTERNAL = (
"ASCEND_LAUNCH_BLOCKING",
"ASCEND_RT_VISIBLE_DEVICES",
"FLASHINFER_DISABLE_VERSION_CHECK",
"HF_HOME",
"HF_HUB_DISABLE_XET",
"HF_TOKEN",
"HIP_VISIBLE_DEVICES",
"HUGGINGFACE_HUB_TOKEN",
"MASTER_ADDR",
"MASTER_PORT",
"MINDIE_SD_FA_TYPE",
"MORI_RDMA_DEVICES",
"MUSA_VISIBLE_DEVICES",
"MV2_COMM_WORLD_LOCAL_RANK",
"NCCL_ASYNC_ERROR_HANDLING",
"OMPI_COMM_WORLD_LOCAL_RANK",
"OPENAI_API_KEY",
"PYTHONPATH",
"QUACK_CACHE_DIR",
"RANK",
"RANK_ID",
"RAY_RAYLET_PID",
"RDMA_DEVICE_NAME",
"SAGE_ATTN_TENSOR_LAYOUT",
"VLLM_LOCAL_RANK",
"WORLD_SIZE",
"XDG_CACHE_HOME",
)
_MODEL_PROMOTE = (
"AR_DIFFUSION_KV_SCRATCH_BLOCKS_PER_BRANCH",
"COSMOS3_SOUND_TOKENIZER_CONFIG_PATH",
"COSMOS3_SOUND_TOKENIZER_PATH",
"COSYVOICE3_ESTIMATOR_ONNX",
"COSYVOICE3_TRT",
"COSYVOICE3_TRT_CACHE",
"DYNIN_CONFIG_PATH",
"DYNIN_MAGVIT_REMOTE_CODE_LOCAL_FILES_ONLY",
"DYNIN_MAGVIT_REMOTE_CODE_REPO_ID",
"DYNIN_MAGVIT_REMOTE_CODE_REVISION",
"DYNIN_REMOTE_CODE_LOCAL_FILES_ONLY",
"DYNIN_REMOTE_CODE_REPO_ID",
"DYNIN_REMOTE_CODE_REVISION",
"FASTVIDEO_VSA_SM100A",
"HIGGS_AUDIO_TOKENIZER_PATH",
"INTERNVLA_A1_COSMOS_DECODER_PATH",
"INTERNVLA_A1_COSMOS_DIR",
"INTERNVLA_A1_COSMOS_ENCODER_PATH",
"INTERNVLA_A1_PROCESSOR_DIR",
"MAGI2_DETERMINISTIC",
"MAGI2_FLASH_ATTN_VERSION",
"MIMO_AUDIO_TOKENIZER_CUDA_GRAPH",
"MIMO_AUDIO_TOKENIZER_DEVICE",
"MIMO_AUDIO_TOKENIZER_PATH",
"MING_CFM_CUDAGRAPH",
"MINICPMO_TOKEN2WAV_TRT",
"MINICPMO_TOKEN2WAV_TRT_CACHE",
"MINICPMO_TOKEN2WAV_TRT_DTYPE",
"NEMOTRON_VOICECHAT_LLM_PATH",
"OMNIVOICE_CUDA_GRAPH",
"OMNIVOICE_TF32",
"VLLM_GEPARD_CHUNK_FRAMES",
"VLLM_GEPARD_FIRST_CHUNK_FRAMES",
"VLLM_GEPARD_LOOKBACK_FRAMES",
"VLLM_OMNI_FISH_KVCACHE_ATTN",
"VLLM_OMNI_SANA_WM_STAGE1_TEXT_ENCODER",
"VLLM_OMNI_SENSENOVA_PAGED_DECODE",
)
_MODEL_REQUEST_SCOPE = (
"DOTS_TTS_DIT_NUM_STEPS",
"GR00T_NOISE_SEED",
"MAGI2_NEGATIVE_PROMPT",
"STEP_AUDIO2_DEFAULT_PROMPT_WAV",
"VLLM_GEPARD_END_FADE_MS",
"VLLM_GEPARD_END_SILENCE_MS",
"VLLM_GEPARD_GREEDY",
)
# No audited model variable is currently owned by a supported third-party
# contract. Keep the disposition explicit so future reviews cannot silently
# fold an externally owned name into "promote".
_MODEL_EXTERNAL: tuple[str, ...] = ()
_MODEL_INTERNALIZE = (
"DOTS_TTS_BETA_TRACE",
"DYNIN_S2U_VENDOR_ROOT",
"DZ_PHASE_TIMING",
"MAGI2_ALLOW_UNSUPPORTED_TOPOLOGY",
"MAGI2_ROUTER_BIAS_SOURCE",
"MING_TTS_STAGE1_FINAL_LOG",
"MOSS_TTS_DEBUG_STOP",
"NEMOTRON_VOICECHAT_DEBUG_FUNCTION_TIMELINE",
"VLLM_OMNI_QWEN3_CODE2WAV_BATCH_STATS",
"VLLM_OMNI_QWEN3_CODE2WAV_BATCH_STATS_LOG_EVERY",
"VLLM_OMNI_QWEN3_CODE2WAV_CUDAGRAPH_STATS",
"VLLM_OMNI_QWEN3_CODE2WAV_CUDAGRAPH_STATS_FILE",
"VLLM_OMNI_QWEN3_CODE2WAV_CUDAGRAPH_STATS_LOG_EVERY",
)
_MODEL_DEPRECATE_REMOVE = (
"HIGGS_AUDIO_V2_TOKENIZER_PATH",
"VLLM_OMNI_LINGBOT_ACTION_ROOT",
"VLLM_OMNI_VOXCPM_CODE_PATH",
"XCODEC1_PATH",
"model_stage",
)
_BENCHMARK_TRANSITIONAL = (
"DAILY_OMNI_EXTRACT_MODE",
"DAILY_OMNI_SAVE_EVAL_ITEMS",
"MAX_NUM_FRAMES",
"SEED_TTS_EVAL_DEVICE",
"SEED_TTS_HF_WHISPER_MODEL",
"SEED_TTS_SIM_DEVICE",
"SEED_TTS_SIM_EVAL",
"SEED_TTS_UTMOS_EVAL",
"SEED_TTS_UTMOS_HF_REPO",
"SEED_TTS_UTMOS_JIT_FILE",
"SEED_TTS_WAVLM_MAX_SECONDS",
"SEED_TTS_WAVLM_MIN_SAMPLES",
"SEED_TTS_WAVLM_MODEL",
"SEED_TTS_WER_EVAL",
"SEED_TTS_WER_SAVE_AUDIO_DIR",
"SEED_TTS_WER_SAVE_ITEMS",
"VLLM_DAILY_OMNI_MEDIA_REPO",
"VLLM_OMNI_BENCH_AUDIO_CHANNELS",
"VLLM_OMNI_BENCH_AUDIO_CONTINUITY_THRESHOLD_S",
"VLLM_OMNI_BENCH_AUDIO_SAMPLE_RATE",
)
_INTERNAL = (
"VLLM_OMNI_DLO_DP_WAVE_TIMEOUT",
"VLLM_OMNI_REPLICA_ID",
# edge branch: set by the orchestrator for its stage subprocesses
"VLLM_OMNI_STAGE_ID",
"VLLM_OMNI_INIT_TIMELINE_DIR",
)
_REDACTED = frozenset(
{
"HF_TOKEN",
"HUGGINGFACE_HUB_TOKEN",
"OPENAI_API_KEY",
}
)
def _build_inventory() -> Mapping[str, EnvironmentVariableClassification]:
inventory: dict[str, EnvironmentVariableClassification] = {}
def add(
names: tuple[str, ...],
category: EnvironmentVariableCategory,
disposition: ModelEnvironmentVariableDisposition | None = None,
) -> None:
for name in names:
if name in inventory:
raise ValueError(f"Environment variable {name!r} is classified more than once")
inventory[name] = EnvironmentVariableClassification(
name=name,
category=category,
model_disposition=disposition,
redact_value=name in _REDACTED,
)
add(_PUBLIC_OMNI, EnvironmentVariableCategory.PUBLIC_OMNI)
add(_INHERITED_VLLM, EnvironmentVariableCategory.INHERITED_VLLM)
add(_PLATFORM_EXTERNAL, EnvironmentVariableCategory.PLATFORM_EXTERNAL)
add(_MODEL_PROMOTE, EnvironmentVariableCategory.MODEL_SPECIFIC, ModelEnvironmentVariableDisposition.PROMOTE)
add(
_MODEL_REQUEST_SCOPE,
EnvironmentVariableCategory.MODEL_SPECIFIC,
ModelEnvironmentVariableDisposition.REQUEST_SCOPE,
)
add(_MODEL_EXTERNAL, EnvironmentVariableCategory.MODEL_SPECIFIC, ModelEnvironmentVariableDisposition.EXTERNAL)
add(
_MODEL_INTERNALIZE,
EnvironmentVariableCategory.MODEL_SPECIFIC,
ModelEnvironmentVariableDisposition.INTERNALIZE,
)
add(
_MODEL_DEPRECATE_REMOVE,
EnvironmentVariableCategory.MODEL_SPECIFIC,
ModelEnvironmentVariableDisposition.DEPRECATE_REMOVE,
)
add(_BENCHMARK_TRANSITIONAL, EnvironmentVariableCategory.BENCHMARK_TRANSITIONAL)
add(_INTERNAL, EnvironmentVariableCategory.INTERNAL)
return MappingProxyType(dict(sorted(inventory.items())))
# All reviewed environment-variable names, indexed by exact name. This is
# classification metadata, not the executable value registry used by
# ``vllm.envs``.
ENVIRONMENT_VARIABLE_INVENTORY: Mapping[str, EnvironmentVariableClassification] = _build_inventory()
Reference in new issue View Git Blame Copy Permalink