File size: 10,724 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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
"""Padding-aligned GRPO trainer for the pinned TRL 1.9.1 + Qwen3.5-VL stack.

TRL 1.9.1's ``GRPOTrainer._generate_and_score_completions`` builds the
multimodal forward inputs for a VLM batch from *two* separate processor calls
that pad on opposite sides:

* ``prompt_ids`` come from ``_tokenize_prompts`` and are later left-padded
  (``padding_side="left"``) so the causal model can decode from the right edge.
* ``mm_token_type_ids`` and ``image_grid_thw`` come from a second call
  ``processing_class(images=..., text=..., padding=True)`` whose Qwen processor
  pads on the **right**.

For a batch whose prompts have unequal token lengths the two tensors no longer
describe the same columns: the image markers in ``mm_token_type_ids`` sit at
different positions than the real image-pad tokens in the left-padded
``prompt_completion_ids``.  ``Qwen3_5VLForConditionalGeneration.get_rope_index``
then groups the (misaligned) ``mm_token_type_ids``, consumes a full
``image_grid_thw`` entry for a *partial* image-marker run, and produces an
``llm_positions`` tensor whose column count exceeds the non-padded token count
of the row — raising::

    RuntimeError: shape mismatch: value tensor of shape [3, 929] cannot be
    broadcast to indexing result of shape [3, 613]

The crash is data- and batch-composition-dependent: equal-length batches pad
trivially and run; a heterogeneous batch (different image patch counts / prompt
lengths) misaligns and crashes.  The same misalignment can also fail *silently*,
dropping an image's vision positions entirely (text-only fallback) without
raising.

The fix recomputes ``mm_token_type_ids`` from the actual ``input_ids`` that
reach the model: every image-pad token marks a vision position, every other
token marks a text position.  This is exactly the construction TRL already uses
for the tool-image path (``grpo_trainer.py`` ~line 2524), ported to the
non-tool path.  Because ``input_ids`` is the left-padded
``prompt_completion_ids`` itself, the rebuilt marker tensor is aligned to it by
construction, regardless of how the second processor call padded.  Only the
per-token marker tensor is rewritten; ``pixel_values``, ``image_grid_thw``,
``num_images``, prompts, completions, masks, and the rollout are untouched, so
vLLM generation (which uses its own path) and all sampling are unchanged — this
is a correctness fix to the training forward, not a hyperparameter.

The override lives on a single shared base so every controlled RL arm
(``grpo``, ``papo_controlled``, ``evi_po``) inherits it.
"""

from __future__ import annotations

import logging
import os
from typing import Any, cast

import torch
from trl import GRPOTrainer

from .vllm_runtime import (
    CUDAGRAPH_ENV_VAR,
    cudagraphs_enabled,
    patched_rollout_engine,
    rollout_engine_kwargs,
)

logger = logging.getLogger(__name__)


def install_reload_weights_skip(vllm_generation: Any) -> bool:
    """No-op vLLM's per-step ``reload_weights`` disk-reload on CUDA colocate.

    TRL 1.9.1's colocate + sleep_mode generate path calls
    ``self.llm.collective_rpc("reload_weights")`` every step as a workaround
    for vLLM issue #29341 (weights going stale after ``sleep``/``wake_up``).
    On CUDA + vLLM 0.25.1 this disk reload leaks file/mmap handles: the
    per-step "Loading safetensors checkpoint shards" duration grows
    exponentially (~0.7s -> ~12s over ~30 steps around step 2900) and
    hard-hangs after ~2900 sleep/wake cycles, which is fatal for the 5750-step
    main/ablation runs (the stall is deterministic -- every retry replays the
    same batch and re-trips the leak).

    The reload is redundant here. ``sync_weights`` already pushes the merged
    LoRA weights into vLLM every step via ``load_weights``
    (``_push_param_to_vllm``), and ``wake_up(tags=["weights"])`` restores the
    weight-memory mapping. That push+wake path is exactly what TRL falls back
    to on backends that do not implement ``reload_weights`` (the
    ``except NotImplementedError`` branch in ``generate``), so skipping the
    disk reload on CUDA is equivalent to that already-supported fallback and
    keeps the rolled-out weights correct every step.

    ``collective_rpc`` is invoked exactly once in TRL's vLLM generation module
    -- with ``"reload_weights"`` -- so narrowing the no-op to that method name
    leaves every other engine RPC untouched. Returns whether the patch was
    applied (the LLM object is absent in non-vLLM / server-mode arms).
    """
    llm = getattr(vllm_generation, "llm", None)
    if llm is None or not callable(getattr(llm, "collective_rpc", None)):
        return False
    original = llm.collective_rpc

    def collective_rpc(method: str, *args: Any, **kwargs: Any) -> Any:
        if method == "reload_weights":
            return None
        return original(method, *args, **kwargs)

    try:
        llm.collective_rpc = collective_rpc  # type: ignore[method-assign]
    except AttributeError:
        logger.warning(
            "vLLM LLM object rejects instance-level collective_rpc override "
            "(__slots__); per-step reload_weights leak NOT patched."
        )
        return False
    logger.info(
        "Patched vLLM collective_rpc to skip per-step reload_weights "
        "(vLLM #29341 workaround leaks on CUDA; sync_weights push suffices)."
    )
    return True


def realign_mm_token_type_ids(
    input_ids: torch.Tensor,
    mm_token_type_ids: torch.Tensor | None,
    image_pad_token_id: int | None,
) -> torch.Tensor | None:
    """Rebuild ``mm_token_type_ids`` aligned to ``input_ids``.

    Returns ``None`` when there is nothing to realign (no multimodal markers, or
    the image-pad token id is unknown). Otherwise returns a new ``long`` tensor
    the same shape as ``input_ids`` with ``1`` at every image-pad token position
    and ``0`` elsewhere — aligned to the (left-padded) ``input_ids`` the model
    sees, instead of the right-padded markers TRL's second processor call emits.
    """
    if mm_token_type_ids is None or image_pad_token_id is None:
        return mm_token_type_ids
    realigned = torch.zeros_like(input_ids, dtype=torch.long)
    realigned[input_ids == image_pad_token_id] = 1
    return realigned


class AlignedGRPOTrainer(GRPOTrainer):  # type: ignore[misc]
    """``GRPOTrainer`` with Qwen3.5-VL ``mm_token_type_ids`` realigned to ``input_ids``."""

    def __init__(self, *args: Any, **kwargs: Any) -> None:
        # Build the vLLM rollout engine with cuda-graphs off. vLLM V1 cuda-graph
        # execution of the Qwen3.5-VL Gated-DeltaNet decode wedges mid-step
        # (GPU 100%, CPU idle, SIGTERM ignored) about once per 600-900 steps,
        # which forces the stall watchdog to SIGKILL and resume. The override
        # must wrap super().__init__ because that is where TRL constructs the
        # engine; see vllm_runtime for why cudagraph_mode beats enforce_eager
        # and why this deliberately stays out of the frozen vllm_config.
        engine_kwargs = rollout_engine_kwargs(os.environ)
        with patched_rollout_engine(engine_kwargs) as patched:
            if patched:
                logger.info(
                    "vLLM rollout engine: cuda-graphs disabled "
                    "(compilation_config=%s) to avoid the V1 cuda-graph "
                    "generation deadlock; set %s=1 to restore upstream behaviour.",
                    engine_kwargs.get("compilation_config"),
                    CUDAGRAPH_ENV_VAR,
                )
            elif cudagraphs_enabled(os.environ):
                logger.info(
                    "vLLM rollout engine: cuda-graphs left ENABLED by %s; "
                    "the V1 cuda-graph generation deadlock can recur "
                    "(~1 per 600-900 steps).",
                    CUDAGRAPH_ENV_VAR,
                )
            super().__init__(*args, **kwargs)
        # Neutralize the per-step vLLM reload_weights disk-reload, which leaks
        # on CUDA and hard-hangs long RL runs after ~2900 sleep/wake cycles.
        # Applied after super().__init__ so ``self.vllm_generation`` exists; it
        # is a per-rank instance, so every DDP worker patches its own vLLM.
        vllm_generation = getattr(self, "vllm_generation", None)
        if vllm_generation is not None:
            install_reload_weights_skip(vllm_generation)

    def _get_per_token_logps_and_entropies(
        self,
        model: Any,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor,
        logits_to_keep: int,
        batch_size: int | None = None,
        compute_entropy: bool = False,
        compute_aux_loss: bool = False,
        pixel_values: torch.Tensor | None = None,
        image_grid_thw: torch.Tensor | None = None,
        num_images: list[int] | None = None,
        pixel_attention_mask: torch.Tensor | None = None,
        spatial_shapes: torch.Tensor | None = None,
        num_tiles: list[int] | None = None,
        image_sizes: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        mm_token_type_ids: torch.Tensor | None = None,
        image_position_ids: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]:
        # Realign the multimodal token-type markers to the (left-padded)
        # input_ids the model is about to see. TRL's non-tool VLM path builds
        # mm_token_type_ids from a second, right-padded processor call, which
        # misaligns from input_ids in heterogeneous-length batches and crashes
        # Qwen3.5-VL's 3D-RoPE index construction. Rebuilding from the image-pad
        # token positions in input_ids restores per-column alignment.
        mm_token_type_ids = realign_mm_token_type_ids(
            input_ids, mm_token_type_ids, getattr(self, "_image_pad_token_id", None)
        )
        return cast(
            tuple[torch.Tensor, torch.Tensor, torch.Tensor | None],
            super()._get_per_token_logps_and_entropies(
                model,
                input_ids,
                attention_mask,
                logits_to_keep,
                batch_size=batch_size,
                compute_entropy=compute_entropy,
                compute_aux_loss=compute_aux_loss,
                pixel_values=pixel_values,
                image_grid_thw=image_grid_thw,
                num_images=num_images,
                pixel_attention_mask=pixel_attention_mask,
                spatial_shapes=spatial_shapes,
                num_tiles=num_tiles,
                image_sizes=image_sizes,
                token_type_ids=token_type_ids,
                mm_token_type_ids=mm_token_type_ids,
                image_position_ids=image_position_ids,
            ),
        )