[EPD][Model Loader] Skip language-model checkpoint shards for --mm-encoder-only (#58086)

Signed-off-by: grYe99 <guorongye99@gmail.com>
Co-authored-by: grYe99 <guorongye99@gmail.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
This commit is contained in:
GuorongYe
2026-09-24 00:41:15 -07:00
committed by GitHub
co-authored by grYe99 Cursor Agent
parent c3f5270270
commit 1291bdb85e
5 changed files with 483 additions and 14 deletions
@@ -0,0 +1,243 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Tests for --mm-encoder-only safetensors shard filtering."""
import json
import os
import tempfile
from transformers.utils import SAFE_WEIGHTS_INDEX_NAME
from vllm.model_executor.model_loader.weight_utils import (
filter_mm_encoder_only_safetensors_files,
resolve_mm_encoder_only_lm_prefixes,
)
from vllm.model_executor.models.utils import WeightsMapper
def _write_index(folder: str, weight_map: dict[str, str]) -> None:
with open(os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME), "w") as f:
json.dump({"weight_map": weight_map}, f)
def _qwen25_mapper() -> WeightsMapper:
"""Qwen2.5-VL dual layout mapper (nested + flat catch-all)."""
return WeightsMapper(
orig_to_new_prefix={
"model.language_model.": "language_model.model.",
"model.visual.": "visual.",
"lm_head.": "language_model.lm_head.",
"model.": "language_model.model.",
}
)
def _qwen3_mapper() -> WeightsMapper:
return WeightsMapper(
orig_to_new_prefix={
"model.visual.": "visual.",
"lm_head.": "language_model.lm_head.",
"model.language_model.": "language_model.model.",
}
)
def _molmo_mapper() -> WeightsMapper:
return WeightsMapper(
orig_to_new_prefix={
"model.vision_backbone.": "vision_backbone.",
"model.transformer.blocks.": "model.layers.",
"model.transformer.ln_f.": "model.norm.",
"model.transformer.mlp.down_proj.": "lm_head.",
}
)
def test_keeps_vision_shards_drops_language_only():
with tempfile.TemporaryDirectory() as folder:
files = [
os.path.join(folder, "model-00001-of-000003.safetensors"),
os.path.join(folder, "model-00002-of-000003.safetensors"),
os.path.join(folder, "model-00003-of-000003.safetensors"),
]
for path in files:
open(path, "wb").close()
_write_index(
folder,
{
"language_model.layers.0.weight": "model-00001-of-000003.safetensors",
"language_model.layers.1.weight": "model-00002-of-000003.safetensors",
"vision_tower.blocks.0.weight": "model-00003-of-000003.safetensors",
"mm_projector.weight": "model-00003-of-000003.safetensors",
},
)
kept = filter_mm_encoder_only_safetensors_files(
files,
folder,
SAFE_WEIGHTS_INDEX_NAME,
("language_model.",),
)
assert kept == [files[2]]
def test_keeps_mixed_shard_with_vision_and_lm():
with tempfile.TemporaryDirectory() as folder:
mixed = os.path.join(folder, "model-00001-of-000001.safetensors")
open(mixed, "wb").close()
_write_index(
folder,
{
"language_model.embed.weight": "model-00001-of-000001.safetensors",
"vision_tower.patch.weight": "model-00001-of-000001.safetensors",
},
)
kept = filter_mm_encoder_only_safetensors_files(
[mixed],
folder,
SAFE_WEIGHTS_INDEX_NAME,
("language_model.",),
)
assert kept == [mixed]
def test_no_index_returns_unchanged():
with tempfile.TemporaryDirectory() as folder:
files = [os.path.join(folder, "model.safetensors")]
open(files[0], "wb").close()
kept = filter_mm_encoder_only_safetensors_files(
files,
folder,
SAFE_WEIGHTS_INDEX_NAME,
("language_model.",),
)
assert kept == files
def test_resolve_returns_module_prefixes_without_hf_hardcode():
"""No global Qwen HF nest list — just vLLM module prefixes."""
prefixes = resolve_mm_encoder_only_lm_prefixes(["language_model"])
assert prefixes == ("language_model.",)
# With Qwen mapper, still module prefixes (keys classified via mapper).
assert resolve_mm_encoder_only_lm_prefixes(
["language_model"], weights_mapper=_qwen25_mapper()
) == ("language_model.",)
def test_resolve_empty_language_model_names_fail_closed():
"""Missing/empty ``_language_model_names`` → None (no language_model. default)."""
assert resolve_mm_encoder_only_lm_prefixes(None) is None
assert resolve_mm_encoder_only_lm_prefixes([]) is None
assert resolve_mm_encoder_only_lm_prefixes(()) is None
def test_resolve_skips_filter_for_shared_hf_root():
"""Shared HF root (Molmo-shaped mapper) → fail-closed, not magic name list."""
assert (
resolve_mm_encoder_only_lm_prefixes(["model"], weights_mapper=_molmo_mapper())
is None
)
# Without a mapper, identity layout: module prefix is returned as-is.
assert resolve_mm_encoder_only_lm_prefixes(["model"]) == ("model.",)
assert resolve_mm_encoder_only_lm_prefixes(["model."]) == ("model.",)
def test_resolve_keeps_llm_prefix_without_model_broadening():
prefixes = resolve_mm_encoder_only_lm_prefixes(["llm"])
assert prefixes == ("llm.",)
def test_qwen3_nested_index_skips_lm_keeps_visual():
with tempfile.TemporaryDirectory() as folder:
files = [
os.path.join(folder, "model-00001-of-000002.safetensors"),
os.path.join(folder, "model-00002-of-000002.safetensors"),
]
for path in files:
open(path, "wb").close()
_write_index(
folder,
{
"model.language_model.layers.0.weight": (
"model-00001-of-000002.safetensors"
),
"lm_head.weight": "model-00001-of-000002.safetensors",
"model.visual.blocks.0.weight": "model-00002-of-000002.safetensors",
},
)
prefixes = resolve_mm_encoder_only_lm_prefixes(
["language_model"], weights_mapper=_qwen3_mapper()
)
assert prefixes is not None
kept = filter_mm_encoder_only_safetensors_files(
files,
folder,
SAFE_WEIGHTS_INDEX_NAME,
prefixes,
weights_mapper=_qwen3_mapper(),
)
assert kept == [files[1]]
def test_qwen25_flat_index_skips_lm_keeps_visual():
with tempfile.TemporaryDirectory() as folder:
files = [
os.path.join(folder, "model-00001-of-000002.safetensors"),
os.path.join(folder, "model-00002-of-000002.safetensors"),
]
for path in files:
open(path, "wb").close()
_write_index(
folder,
{
"model.layers.0.self_attn.q_proj.weight": (
"model-00001-of-000002.safetensors"
),
"model.embed_tokens.weight": "model-00001-of-000002.safetensors",
"model.norm.weight": "model-00001-of-000002.safetensors",
"lm_head.weight": "model-00001-of-000002.safetensors",
"visual.blocks.0.weight": "model-00002-of-000002.safetensors",
},
)
prefixes = resolve_mm_encoder_only_lm_prefixes(
["language_model"], weights_mapper=_qwen25_mapper()
)
assert prefixes is not None
kept = filter_mm_encoder_only_safetensors_files(
files,
folder,
SAFE_WEIGHTS_INDEX_NAME,
prefixes,
weights_mapper=_qwen25_mapper(),
)
assert kept == [files[1]]
def test_molmo_shaped_index_not_filtered_when_deny_disabled():
"""Shared-root LM attr → resolve None; callers leave the file list alone."""
assert (
resolve_mm_encoder_only_lm_prefixes(["model"], weights_mapper=_molmo_mapper())
is None
)
with tempfile.TemporaryDirectory() as folder:
files = [
os.path.join(folder, "vision.safetensors"),
os.path.join(folder, "lm.safetensors"),
]
for path in files:
open(path, "wb").close()
_write_index(
folder,
{
"model.vision_backbone.patch.weight": "vision.safetensors",
"model.transformer.blocks.0.weight": "lm.safetensors",
},
)
# Document why raw HF startswith("model.") is unsafe without fail-closed:
# whole-shard deny would also drop vision.
if_applied_naively = filter_mm_encoder_only_safetensors_files(
files, folder, SAFE_WEIGHTS_INDEX_NAME, ("model.",)
)
assert if_applied_naively == []
@@ -79,7 +79,7 @@ def test_default_loader_hf_still_falls_back_to_pt(tmp_path):
# Control: load_format="hf" still picks up .pt weights via fallback.
(tmp_path / "model.pt").write_bytes(b"\x00\x00\x00\x00")
loader = DefaultModelLoader(LoadConfig(load_format="hf"))
_, files, use_safetensors = loader._prepare_weights(
_, files, use_safetensors, _ = loader._prepare_weights(
str(tmp_path),
None,
None,
@@ -1,11 +1,13 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import dataclasses
import glob
import os
import time
from collections.abc import Generator, Iterable
from typing import cast
from typing import TYPE_CHECKING, cast
import torch
from torch import nn
@@ -25,6 +27,7 @@ from vllm.model_executor.model_loader.weight_utils import (
fastsafetensors_weights_iterator,
filter_duplicate_safetensors_files,
filter_files_not_needed_for_inference,
filter_mm_encoder_only_safetensors_files,
get_quant_config,
instanttensor_weights_iterator,
maybe_download_from_modelscope,
@@ -32,11 +35,15 @@ from vllm.model_executor.model_loader.weight_utils import (
multi_thread_safetensors_weights_iterator,
np_cache_weights_iterator,
pt_weights_iterator,
resolve_mm_encoder_only_lm_prefixes,
safetensors_weights_iterator,
)
from vllm.tracing import instrument
from vllm.transformers_utils.repo_utils import list_filtered_repo_files
if TYPE_CHECKING:
from vllm.model_executor.models.utils import WeightsMapper
logger = init_logger(__name__)
@@ -74,6 +81,9 @@ class DefaultModelLoader(BaseModelLoader):
def __init__(self, load_config: LoadConfig):
super().__init__(load_config)
self.local_expert_ids: set[int] | None = None
# Set in load_weights when --mm-encoder-only; used to drop LM-only shards.
self._encoder_only_lm_prefixes: tuple[str, ...] | None = None
self._encoder_only_weights_mapper: WeightsMapper | None = None
extra_config = load_config.model_loader_extra_config
if not isinstance(extra_config, dict):
@@ -132,7 +142,7 @@ class DefaultModelLoader(BaseModelLoader):
revision: str | None,
fall_back_to_pt: bool,
allow_patterns_overrides: list[str] | None,
) -> tuple[str, list[str], bool]:
) -> tuple[str, list[str], bool, str]:
"""Prepare weights for the model.
If the model is not local, it will be downloaded."""
@@ -239,20 +249,40 @@ class DefaultModelLoader(BaseModelLoader):
f"Cannot find any model weights with `{model_name_or_path}`"
)
return hf_folder, hf_weights_files, use_safetensors
return hf_folder, hf_weights_files, use_safetensors, index_file
def _get_weights_iterator(
self, source: "Source"
self, source: Source
) -> Generator[tuple[str, torch.Tensor], None, None]:
"""Get an iterator for the model weights based on the load format."""
extra_config = self.load_config.model_loader_extra_config
hf_folder, hf_weights_files, use_safetensors = self._prepare_weights(
source.model_or_path,
source.subfolder,
source.revision,
source.fall_back_to_pt,
source.allow_patterns_overrides,
hf_folder, hf_weights_files, use_safetensors, index_file = (
self._prepare_weights(
source.model_or_path,
source.subfolder,
source.revision,
source.fall_back_to_pt,
source.allow_patterns_overrides,
)
)
if (
self._encoder_only_lm_prefixes is not None
and use_safetensors
and hf_weights_files
):
hf_weights_files = filter_mm_encoder_only_safetensors_files(
hf_weights_files,
hf_folder,
index_file,
self._encoder_only_lm_prefixes,
weights_mapper=self._encoder_only_weights_mapper,
)
if not hf_weights_files:
raise RuntimeError(
"mm-encoder-only shard filter removed all weight files for "
f"`{source.model_or_path}`; check language_model prefixes "
f"{self._encoder_only_lm_prefixes}"
)
if self.load_config.load_format == "npcache":
# Currently np_cache only support *.bin checkpoints
assert use_safetensors is False
@@ -348,6 +378,29 @@ class DefaultModelLoader(BaseModelLoader):
allow_patterns_overrides=None,
)
def _init_mm_encoder_only_weight_filter(
self, model: nn.Module, model_config: ModelConfig
) -> None:
"""Skip pure language-model safetensors shards under --mm-encoder-only."""
mm_config = model_config.multimodal_config
if mm_config is None or not mm_config.mm_encoder_only:
self._encoder_only_lm_prefixes = None
self._encoder_only_weights_mapper = None
return
# Derive from _language_model_names; fail-closed on shared HF roots
# (Molmo/Phi-4-MM/Muse). Qwen nested/flat keys classified via mapper.
weights_mapper = cast(
"WeightsMapper | None", getattr(model, "hf_to_vllm_mapper", None)
)
self._encoder_only_lm_prefixes = resolve_mm_encoder_only_lm_prefixes(
getattr(model, "_language_model_names", None),
weights_mapper=weights_mapper,
)
self._encoder_only_weights_mapper = (
weights_mapper if self._encoder_only_lm_prefixes is not None else None
)
def _init_ep_weight_filter(self, model_config: ModelConfig) -> None:
"""Compute local expert ids for EP weight filtering.
@@ -423,6 +476,7 @@ class DefaultModelLoader(BaseModelLoader):
self.load_config.safetensors_load_strategy = "torchao"
self._init_ep_weight_filter(model_config)
self._init_mm_encoder_only_weight_filter(model, model_config)
loaded_weights = model.load_weights(self.get_all_weights(model_config, model))
@@ -15,7 +15,7 @@ import time
from collections.abc import Callable, Generator, Iterable
from contextlib import contextmanager
from pathlib import Path
from typing import IO, Any
from typing import IO, TYPE_CHECKING, Any
import filelock
import huggingface_hub.constants
@@ -61,6 +61,9 @@ except ImportError:
from vllm.model_executor.layers.quantization.torchao import torchao_version_at_least
if TYPE_CHECKING:
from vllm.model_executor.models.utils import WeightsMapper
logger = init_logger(__name__)
# use system-level temp directory for file locks, so that multiple users
@@ -605,6 +608,171 @@ def filter_files_not_needed_for_inference(hf_weights_files: list[str]) -> list[s
return hf_weights_files
def _normalize_module_prefixes(
language_model_names: Iterable[str] | None,
) -> tuple[str, ...]:
names = tuple(language_model_names or ())
return tuple(n if n.endswith(".") else f"{n}." for n in names)
def _mapped_weight_name(weights_mapper: "WeightsMapper | None", key: str) -> str | None:
"""Apply ``WeightsMapper.map_name`` when present; else identity."""
if weights_mapper is None:
return key
return weights_mapper.map_name(key)
def _hf_prefix_maps_to_lm(
new_name: str | None, lm_module_prefixes: tuple[str, ...]
) -> bool:
if not new_name:
return False
new_p = new_name if new_name.endswith(".") else f"{new_name}."
return any(new_p.startswith(p) or p.startswith(new_p) for p in lm_module_prefixes)
def _shared_hf_root_module_prefixes(
lm_module_prefixes: tuple[str, ...],
weights_mapper: "WeightsMapper | None",
) -> tuple[str, ...]:
"""Module prefixes that are also an HF root for non-LM weights.
E.g. Molmo/Phi-4-MM/Muse mark LM as ``model`` while vision HF keys stay
under ``model.vision_*`` / ``model.embed_tokens_extend.*``. Detected via
``WeightsMapper.orig_to_new_prefix`` (no hard-coded name list).
"""
if weights_mapper is None:
return ()
mapping = weights_mapper.orig_to_new_prefix
if not mapping:
return ()
non_lm_origs: list[str] = []
for orig, new in mapping.items():
orig_p = orig if orig.endswith(".") else f"{orig}."
if not _hf_prefix_maps_to_lm(new, lm_module_prefixes):
non_lm_origs.append(orig_p)
if not non_lm_origs:
return ()
shared = [
p
for p in lm_module_prefixes
if any(orig.startswith(p) for orig in non_lm_origs)
]
return tuple(shared)
def resolve_mm_encoder_only_lm_prefixes(
language_model_names: Iterable[str] | None,
*,
weights_mapper: "WeightsMapper | None" = None,
) -> tuple[str, ...] | None:
"""Resolve vLLM LM *module* prefixes for ``--mm-encoder-only`` shard skip.
Prefixes come from ``_language_model_names``. Classification of HF index
keys is done later via optional ``weights_mapper`` (see
``filter_mm_encoder_only_safetensors_files``), so this helper does **not**
hard-code Qwen/HF nest lists.
Returns ``None`` (leave the safetensors file list unchanged) when
``_language_model_names`` is empty/missing, or when a module prefix is a
shared HF checkpoint root for both LM and non-LM weights — fail-closed so
Molmo / Phi-4-MM / Muse cannot under-load the encoder.
"""
prefixes = _normalize_module_prefixes(language_model_names)
if not prefixes:
return None
shared = _shared_hf_root_module_prefixes(prefixes, weights_mapper)
if shared:
logger.warning_once(
"mm-encoder-only whole-shard filter not applied: LM module "
"prefix(es) %s share an HF checkpoint root with non-LM weights "
"(via WeightsMapper). Encoder-only load keeps the full "
"safetensors file list; finer HF prefixes or key-level "
"filtering can enable skip later.",
shared,
)
return None
return prefixes
def filter_mm_encoder_only_safetensors_files(
hf_weights_files: list[str],
hf_folder: str,
index_file: str,
language_model_prefixes: Iterable[str],
*,
weights_mapper: "WeightsMapper | None" = None,
) -> list[str]:
"""Drop safetensors shards that only contain language-model weights.
Used with ``--mm-encoder-only`` so Encoder-only EPD instances avoid reading
pure LM shards from disk/DRAM.
Each HF index key is classified by mapping through ``weights_mapper`` (when
provided) and testing the *vLLM* name against ``language_model_prefixes``.
Without a mapper, HF names are compared directly (identity checkpoint
layout, e.g. Kimi ``language_model.*``).
A shard is kept if it contains any non-LM key. Without an index file,
returns ``hf_weights_files`` unchanged.
"""
prefixes = tuple(language_model_prefixes)
if not prefixes:
return hf_weights_files
index_path = os.path.join(hf_folder, index_file)
if not os.path.isfile(index_path):
logger.warning_once(
"mm-encoder-only shard filter skipped: index file %s not found",
index_path,
)
return hf_weights_files
with open(index_path) as f:
weight_map: dict[str, str] = json.load(f)["weight_map"]
keys_by_file: dict[str, list[str]] = {}
for weight_name, weight_file in weight_map.items():
keys_by_file.setdefault(weight_file, []).append(weight_name)
def _is_lm_key(key: str) -> bool:
mapped = _mapped_weight_name(weights_mapper, key)
if mapped is None:
return False
return any(mapped.startswith(p) for p in prefixes)
def _is_lm_only(weight_file: str) -> bool:
keys = keys_by_file.get(weight_file)
if not keys:
# Not listed in the index (e.g. extra matched glob) — keep.
return False
return all(_is_lm_key(key) for key in keys)
kept: list[str] = []
skipped = 0
for path in hf_weights_files:
rel = os.path.relpath(path, hf_folder)
# Index weight_map values are usually basenames; tolerate either form.
basename = os.path.basename(path)
if _is_lm_only(rel) or _is_lm_only(basename):
skipped += 1
continue
kept.append(path)
if skipped:
logger.info_once(
"mm-encoder-only: skipped %d/%d safetensors shard(s) that only "
"contain language-model weights (prefixes=%s)",
skipped,
skipped + len(kept),
prefixes,
)
return kept
# explicitly use pure text format, with a newline at the end
# this makes it impossible to see the animation in the progress bar
# but will avoid messing up with ray or multiprocessing, which wraps
+6 -2
View File
@@ -78,11 +78,15 @@ class WeightsMapper:
orig_to_new_suffix={**self.orig_to_new_suffix, **other.orig_to_new_suffix},
)
def _map_name(self, key: str) -> str | None:
"""Map a weight name (backward-compatible wrapper that discards shard_id)."""
def map_name(self, key: str) -> str | None:
"""Map a weight name; returns ``None`` if the weight should be ignored."""
result = self._map_name_with_shard(key)
return result[0] if result is not None else None
def _map_name(self, key: str) -> str | None:
"""Backward-compatible alias for :meth:`map_name`."""
return self.map_name(key)
def _map_name_with_shard(self, key: str) -> tuple[str, ShardIds | None] | None:
"""Map a weight name and extract any shard_id metadata.