File size: 6,794 Bytes
e1ced61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
"""Rollout-engine overrides for the pinned TRL 1.9.1 + vLLM 0.25.1 stack.

TRL 1.9.1 builds the colocate rollout engine from a fixed ``LLM(...)`` kwarg
list in ``trl/generation/vllm_generation.py`` and exposes **no** compilation or
cuda-graph knob: ``GRPOConfig`` carries ``vllm_mode``,
``vllm_gpu_memory_utilization``, ``vllm_max_model_length``,
``vllm_tensor_parallel_size``, ``vllm_enable_sleep_mode``, ``vllm_model_impl``
and ``vllm_structured_outputs_regex``, and nothing for ``enforce_eager`` or
``compilation_config``. vLLM 0.25.1 has no ``VLLM_ENFORCE_EAGER`` environment
escape hatch either (``vllm/envs.py``). On this pinned stack the only way to
change the rollout engine's cuda-graph behaviour is therefore to inject the
kwarg at TRL's ``LLM(...)`` call site.

Why we inject it: vLLM V1 cuda-graph execution intermittently wedges mid-step
during GRPO generation (GPU pinned at 100%, CPU idle, no traceback, SIGTERM
ignored), roughly once per 600-900 optimizer steps, and clears on retry from the
same checkpoint. Every ``py-spy`` frame collected across occurrences lands in
cuda-graph execution of the Qwen3.5-VL Gated-DeltaNet text backbone
(``vllm/compilation/cuda_graph.py`` ``execution_fn``,
``vllm/model_executor/models/qwen3_next.py`` ``Qwen3NextModel.forward``, the
``rearrange_mixed_qkv`` / ``qwen_gdn_attention_core`` GDN kernels) or in the
stream ``synchronize`` that waits on it (``gpu_model_runner.get_output``). The
hang is not output-driven: completions at the stalled steps are short
(mean 8-12, max <=24 tokens) with ``clipped_ratio`` 0.

``cudagraph_mode="NONE"`` is preferred over ``enforce_eager=True``. Both remove
graph capture and replay -- the implicated mechanism -- but ``enforce_eager``
*additionally* disables ``torch.compile``, giving up the Inductor-compiled
kernels for the whole rollout. ``cudagraph_mode`` is an independent field from
``mode`` (``CompilationMode``) in vLLM 0.25.1's ``CompilationConfig``, so
``NONE`` keeps compilation and drops only the graph layer: the strictly smaller
change, and the cheaper one for a rollout whose decode is only ~12 tokens deep
(cuda-graphs mainly amortise launch overhead across many decode steps, so little
is being given up here, while a stall costs the watchdog's 120-180s detection
window plus a process restart and engine re-init).

Consequences to keep in mind:

* This does not go through ``vllm_config``, which is a frozen smoke-gate key
  (``smoke_gate.FROZEN_COMMON_KEYS``). The frozen common contract stays
  byte-identical, so this needs only a re-freeze at the new commit, not a
  re-smoke -- the same reasoning as the ``reload_weights`` skip in ``eafc6a5``.
* Sampling is unchanged in intent (seed, prompt, sampling params, batch sizes
  are untouched), but disabling graph capture removes vLLM's padding of decode
  batches up to captured graph sizes, so reduction shapes can differ and
  completions are not guaranteed bit-identical to a cuda-graph run. Introduce it
  at an arm boundary rather than mid-arm when that matters.
* Because the knob is not part of ``vllm_config``, the run manifest cannot
  distinguish the two settings; the applied state is logged at ``INFO`` into
  ``queue.log`` instead, and is pinned by ``code_commit`` for the default.

Set ``EXPLICIT_VLLM_CUDAGRAPHS=1`` to restore upstream cuda-graph behaviour
(accepting the hang risk), which is what an A/B measurement of the throughput
cost or a confirmation of the root cause wants.

This module deliberately imports no ``torch`` / ``trl`` / ``vllm`` so the policy
stays unit-testable in the CPU test venv; ``aligned_grpo`` applies it.
"""

from __future__ import annotations

import contextlib
import importlib
import logging
from collections.abc import Iterator, Mapping
from typing import Any

logger = logging.getLogger(__name__)

CUDAGRAPH_ENV_VAR = "EXPLICIT_VLLM_CUDAGRAPHS"

TRL_VLLM_MODULE = "trl.generation.vllm_generation"

_TRUTHY = frozenset({"1", "true", "yes", "on"})

# vLLM 0.25.1 ``CompilationConfig.cudagraph_mode``; ``"NONE"`` means no
# cudagraph capture. Passed as a plain dict because ``LLM()`` accepts
# ``int | dict | CompilationConfig`` and validates a string enum name via
# ``CUDAGraphMode[value.upper()]`` -- so we need not import vllm here.
_NO_CUDAGRAPH_COMPILATION_CONFIG: Mapping[str, Any] = {"cudagraph_mode": "NONE"}


def cudagraphs_enabled(env: Mapping[str, str]) -> bool:
    """Whether the rollout engine should keep upstream cuda-graph behaviour.

    Defaults to ``False`` (graphs disabled): unattended long runs must not pay
    the deadlock retry tax. ``EXPLICIT_VLLM_CUDAGRAPHS`` opts back in.
    """
    return (env.get(CUDAGRAPH_ENV_VAR) or "").strip().lower() in _TRUTHY


def rollout_engine_kwargs(env: Mapping[str, str]) -> dict[str, Any]:
    """Extra ``LLM(...)`` kwargs for TRL's colocate rollout engine."""
    if cudagraphs_enabled(env):
        return {}
    return {"compilation_config": dict(_NO_CUDAGRAPH_COMPILATION_CONFIG)}


@contextlib.contextmanager
def patched_rollout_engine(extra_kwargs: Mapping[str, Any]) -> Iterator[bool]:
    """Inject ``extra_kwargs`` into TRL's colocate ``LLM(...)`` construction.

    TRL imports ``LLM`` into its vLLM generation module at import time (guarded
    by ``is_vllm_available()``), so rebinding that module attribute for the
    duration of engine construction is enough to reach the single call site.
    Kwargs TRL passes explicitly always win, so this can only *add* settings
    that TRL leaves at their vLLM default.

    Yields whether the patch was installed, and always restores the original
    binding. A missing module or absent ``LLM`` (vLLM not installed, or TRL
    moved the call site) is a no-op with a warning rather than a hard failure:
    the engine still builds, just without the override.
    """
    if not extra_kwargs:
        yield False
        return
    try:
        module = importlib.import_module(TRL_VLLM_MODULE)
    except ImportError:
        logger.warning(
            "Cannot import %s; vLLM rollout-engine override NOT applied.", TRL_VLLM_MODULE
        )
        yield False
        return
    original = getattr(module, "LLM", None)
    if not callable(original):
        logger.warning(
            "%s exposes no callable LLM; vLLM rollout-engine override NOT applied.",
            TRL_VLLM_MODULE,
        )
        yield False
        return

    def construct_llm(*args: Any, **kwargs: Any) -> Any:
        return original(*args, **{**extra_kwargs, **kwargs})

    # setattr/getattr rather than attribute syntax: the binding only exists when
    # TRL imported vLLM, so it is not part of the module's static surface.
    setattr(module, "LLM", construct_llm)  # noqa: B010
    try:
        yield True
    finally:
        setattr(module, "LLM", original)  # noqa: B010