Compare commits

...

15 Commits

Author SHA1 Message Date
Wing Lian
d17ed89a3c add missing file 2026-04-21 08:44:01 -04:00
Wing Lian
02e4f2350d fixes for scattermoe from latest peft upgrade 2026-04-21 08:00:16 -04:00
Wing Lian
4195605ab2 fix test dims 2026-04-21 00:44:26 +00:00
Wing Lian
37acb28d02 fix einsum dims 2026-04-20 23:09:47 +00:00
Wing Lian
4a5281e61a Fix shape 2026-04-19 01:53:05 +00:00
Wing Lian
a892d8cce1 chore: lint 2026-04-17 17:48:26 +00:00
Wing Lian
78de2919a6 tiled mlp fix for gemma4 2026-04-16 13:24:41 +00:00
Wing Lian
28283ff373 revert shared_kv_states workaround with transformers 5.5.4 2026-04-15 13:32:59 +00:00
Wing Lian
dc16859983 [gemma4] fix fused RMSNorm+RoPE on hybrid attention models
- Kernel: fused_rms_norm_rope crashed when cos.shape[-1] < x.shape[-1].
  Triton forward/backward take an n_rot runtime arg that restricts
  rotate_half to [0, n_rot) and treats trailing cols as RMSNorm-only
  pass-through (cos=1, sin=0 defaults). Wrapper also expands cos/sin
  that broadcast over batch.

- Forward: _make_fused_forward used a stale shared_kv_states kwarg the
  current decoder layer no longer passes. Now mirrors stock attention,
  reading/writing past_key_values.shared_layers.
2026-04-15 13:27:31 +00:00
Wing Lian
d4e9cf2eec lint 2026-04-15 13:27:30 +00:00
Wing Lian
53391a10d7 vllm-serve-lora add /v1/completions route + worker pipe lock
The LoRA vllm-serve wrapper only exposed /v1/chat/completions, but
retrace's SWE agent server uses the token-id-aware /v1/completions
endpoint so it can feed raw prompt_token_ids + track per-token
logprobs across multi-turn rollouts. Add the route, mirroring the
shape of /v1/chat/completions but routing to the vLLM worker's
generate() method so prompt_token_ids are passed through as-is.

Also add a worker_pipe_lock around conn.send/conn.recv. The
multiprocessing.Connection to the vLLM worker is a single shared
full-duplex pipe; concurrent HTTP requests interleave pickle frames
on the wire and corrupt the stream (observed as
UnpicklingError: pickle data was truncated, surfacing as 500s).
The agent server fires ~8 concurrent rollout requests at once, so
this was a hard blocker for any multi-concurrent workload. Serialize
access to the pipe per-request round-trip.
2026-04-15 13:27:30 +00:00
Wing Lian
7617b951a8 make _maybe_sync_vllm_weights actually fire in sync mode
Two bugs in ``AsyncGRPOTrainer._maybe_sync_vllm_weights`` plus a
companion bug in the sync-hook patch site that together neutralized
LoRA weight sync entirely whenever ``async_prefetch=False`` was
combined with NeMo Gym's data-producer path:

1. ``_maybe_sync_vllm_weights`` had ``if not async_prefetch: return``
   at the top. The original design assumed sync mode would fall back
   to TRL's stock per-step ``sync_weights`` call inside
   ``_generate_single_turn`` — true for vanilla GRPO but FALSE in
   NeMo Gym multi-turn, where ``NemoGymDataProducer`` calls the agent
   server directly and ``_generate_single_turn`` is never invoked.
   Result: no sync ever happened in NeMo Gym sync mode.

2. ``step % vllm_sync_interval`` would TypeError on the first call if
   ``vllm_sync_interval`` was unset (the default for any config that
   doesn't explicitly set it).

3. The ``_generate_single_turn`` patch installed
   ``vllm_generation.sync_weights = lambda: None`` unconditionally
   for vllm_lora_sync runs. That's correct in async-prefetch mode
   (BG thread can't safely sync) but wrong in sync mode: TRL's
   per-step auto-sync inside ``_generate_single_turn`` was the
   fallback that the early return in (1) was assuming, and the
   no-op patch was killing it.

Fix:
  - Drop the ``not async_prefetch`` early return; ``_maybe_sync_vllm_weights``
    is now the canonical sync trigger and runs in both modes from
    ``_prepare_inputs_with_data_producer`` / ``_prepare_inputs_legacy_async``.
  - Default ``vllm_sync_interval`` to 1 when unset.
  - In the ``_generate_single_turn`` patch, route sync_weights to
    ``_sync_lora_adapter`` in sync mode (and keep the lambda no-op
    in async mode for the BG-thread safety reason).
2026-04-15 13:27:30 +00:00
Wing Lian
e993ed5208 retry head-server probe with longer timeout
``get_server_configs`` was hardcoded to a 5s timeout with no retry.
That's empirically too tight to survive a kill-and-relaunch cycle:
when the agent server is finishing in-flight rollouts from a prior
run, it can take 10-30s to respond to /global_config_dict_yaml, and
the trainer would crash at startup with a ReadTimeoutError.

Bump the per-attempt timeout to 30s and retry up to 3 times with a
2s/4s backoff. The retry intentionally raises a RuntimeError after
the third failure rather than returning empty config — silent
failure here would let training proceed with no agent servers
discovered, which is also a no-op trainer.
2026-04-15 13:27:30 +00:00
Wing Lian
69f165b39b probe vLLM weight-sync routes and select transport per server
The plugin used to unconditionally monkey-patch
VLLMClient.init_communicator to a no-op AND silently no-op
sync_weights when vllm_lora_sync was off. Combined, this turned the
trainer into a functional no-op whenever (a) the user ran NeMo Gym
+ LoRA without remembering to set vllm_lora_sync=true or (b) the
user ran NeMo Gym + full fine-tune (which had no working sync path
under the old code).

Replace both patches with:

1. A probe of the configured vLLM server's /openapi.json at
   pre_model_load. Three transports are recognized:
     - NCCL (/init_communicator/ + /update_named_param/) — TRL serve
       and axolotl vllm-serve both expose this
     - LoRA filesystem (/v1/load_lora_adapter or /set_lora_adapter/)
     - HTTP base64 full-weight (/http_update_weights/) — axolotl
       vllm-serve only

2. A pure-logic ``select_weight_sync_transport`` that picks the
   right one for (server caps × adapter type).

3. ``init_communicator`` is only patched out when the server has no
   NCCL routes; against TRL/axolotl serve modules it stays live so
   full-finetune NCCL sync works.

4. ``post_trainer_create`` uses the selection table to install LoRA
   filesystem sync OR leave the standard NCCL flow alone OR raise
   NotImplementedError (HTTP — pending) OR raise a precise diagnosis
   when no transport is viable. No more silent no-op trainers.
2026-04-15 13:27:30 +00:00
Wing Lian
80a97f192b validate batch shape against num_generations at config time
Surfaces a class of GRPO config errors at axolotl-train startup instead
of letting them bubble out of GRPOTrainer.__init__ after the model loads.
Three checks under RLValidationMixin.check_grpo_batch_size_divisibility:

  - effective generation_batch_size (or mb*GA fallback) must be divisible
    by trl.num_generations, with a hint pointing at the smallest GA bump
    that fixes the violation
  - num_generations >= 2 (group-relative advantage needs variance; with
    num_gen=1 the policy never updates)
  - When world_size > 1, effective gbs >= num_generations * world_size

11 unit tests cover the table: divisible/non-divisible, explicit and
implicit gbs, multi-rank constraint, GRPO-disabled passthrough, and
unset num_generations.
2026-04-15 13:27:30 +00:00
23 changed files with 2596 additions and 285 deletions

View File

@@ -242,6 +242,85 @@ class ProducerConfig:
)
class _GroupShardedSampler:
"""Rank-aware shard of a ``RepeatSampler`` that preserves GRPO groups.
``RepeatSampler`` yields ``num_generations`` consecutive copies of
each prompt, forming a GRPO group. For distributed training each
rank must see a disjoint slice of prompts (otherwise every rank
dogpiles on the first 1/world_size of the batch) while keeping each
group intact on a single rank so advantage normalization sees all
peer generations.
``accelerator.prepare(DataLoader)`` does not handle this correctly
for custom samplers with ``split_batches=False`` (the default): it
leaves the sampler alone and every rank replays identical indices.
This wrapper fixes that by consuming the inner sampler's full
output, chunking it into ``num_generations``-sized groups, and
round-robining whole groups across ranks.
Intended to be used ONLY when distributed training is active
(``num_replicas > 1``); for single-rank it is a no-op but still
correct.
"""
def __init__(
self,
inner: Any,
num_generations: int,
rank: int,
num_replicas: int,
):
if num_generations < 1:
raise ValueError(f"num_generations must be >= 1, got {num_generations}")
if num_replicas < 1:
raise ValueError(f"num_replicas must be >= 1, got {num_replicas}")
if not (0 <= rank < num_replicas):
raise ValueError(f"rank must be in [0, {num_replicas}), got {rank}")
self.inner = inner
self.num_generations = num_generations
self.rank = rank
self.num_replicas = num_replicas
def __iter__(self):
all_indices = list(self.inner)
if len(all_indices) % self.num_generations != 0:
raise ValueError(
f"inner sampler yielded {len(all_indices)} indices, "
f"not a multiple of num_generations={self.num_generations}"
)
# Chunk the flat index sequence into groups of num_generations
# consecutive indices. ``RepeatSampler`` guarantees that each
# group contains num_generations copies of the same prompt id.
groups = [
all_indices[i : i + self.num_generations]
for i in range(0, len(all_indices), self.num_generations)
]
# Round-robin whole groups across ranks. Round-robin (vs.
# contiguous chunking) preserves approximate shuffled order on
# each rank even when the group count is small relative to the
# world size.
for group in groups[self.rank :: self.num_replicas]:
yield from group
def __len__(self):
try:
inner_len = len(self.inner)
except TypeError:
# Non-sized inner sampler — we can't know the per-rank
# length without materializing. Return 0 as a hint that the
# DataLoader should fall back to iteration.
return 0
total_groups = inner_len // self.num_generations
# Ceiling division for the trailing groups that don't divide
# evenly — extra groups go to the first ``total_groups %
# num_replicas`` ranks, matching the round-robin above.
my_groups = (
total_groups + self.num_replicas - self.rank - 1
) // self.num_replicas
return my_groups * self.num_generations
class DataProducer(ABC):
"""Abstract base class for online data producers.
@@ -556,6 +635,34 @@ class GRPODataProducer(BaseDataProducer):
seed=self._seed,
)
# Shard the sampler across distributed ranks so each rank sees
# a disjoint slice of prompts. ``RepeatSampler`` groups each
# prompt with ``num_generations`` consecutive copies — our
# wrapper round-robins WHOLE groups across ranks so all
# generations of a given prompt stay on the same rank (needed
# for GRPO advantage normalization within a group).
#
# Without this, ``accelerator.prepare(dl)`` with the default
# ``split_batches=False`` leaves the custom sampler alone, so
# every rank iterates the identical index sequence and the
# cluster dogpiles on the first 1/world_size of the prompts.
num_replicas = max(1, trainer.accelerator.num_processes)
if num_replicas > 1:
sampler = _GroupShardedSampler(
inner=sampler,
num_generations=self._num_generations,
rank=trainer.accelerator.process_index,
num_replicas=num_replicas,
)
logger.info(
"[RANK:%d] _GroupShardedSampler active "
"(num_replicas=%d, num_generations=%d, gen_batch=%d)",
trainer.accelerator.process_index,
num_replicas,
self._num_generations,
self._generation_batch_size,
)
# Use identity collator (same as stock GRPOTrainer)
def _identity(x):
return x
@@ -574,12 +681,11 @@ class GRPODataProducer(BaseDataProducer):
rank=trainer.args.process_index,
),
)
self._prompt_dl = trainer.accelerator.prepare(dl)
# Don't let accelerator track this dataloader
acc_dls = trainer.accelerator._dataloaders
if self._prompt_dl in acc_dls:
acc_dls.remove(self._prompt_dl)
# Skip accelerator.prepare — we're handling per-rank sharding
# ourselves via ``_GroupShardedSampler``. ``prepare()`` would
# otherwise try to wrap the DataLoader with its own sharding
# logic which does not understand our group structure.
self._prompt_dl = dl
self._prompt_iter = iter(self._prompt_dl)
@@ -1103,11 +1209,22 @@ class AsyncGRPOTrainer(GRPOTrainer):
- vllm_lora_sync: saves adapter to filesystem, vLLM loads natively
- PEFT no-merge: computes merged weights as new tensors, NCCL broadcast
- Non-PEFT: stock sync_weights via merge_adapter + NCCL
This is the canonical sync trigger and runs in BOTH async and
synchronous modes from ``_prepare_inputs_with_data_producer`` /
``_prepare_inputs_legacy_async``. The ``_generate_single_turn``
patch is a parallel backup for non-data-producer paths (vanilla
GRPO without NeMo Gym), where the data producer is bypassed
entirely and TRL's stock generate-then-sync flow is used instead.
"""
if not (self.use_vllm and self.args.async_prefetch):
if not self.use_vllm:
return
step = self.state.global_step
interval = self.args.vllm_sync_interval
# Default to syncing every step when no interval is configured —
# otherwise ``step % None`` would TypeError, and the previous
# behavior of crashing on the first sync was strictly worse than
# the standard "sync every optimizer step".
interval = self.args.vllm_sync_interval or 1
if step != self._last_synced_step and step % interval == 0:
if step == 0:
logger.info("Skipping vLLM weight sync at step 0 (no training yet)")
@@ -1202,13 +1319,42 @@ class AsyncGRPOTrainer(GRPOTrainer):
# Permanently replace vllm_generation.sync_weights with our custom
# sync to avoid merge_adapter (fails on FP8 / races with training).
# For LoRA sync mode, make it a no-op here since _maybe_sync_vllm_weights
# handles the sync with proper interval tracking.
#
# The design has two modes that have to be threaded carefully:
#
# - Async prefetch ON: BG generation thread can't safely call
# sync_weights mid-rollout (it races with the trainer's optimizer
# step and can corrupt weights). We no-op the stock sync hook and
# drive sync ourselves from ``_maybe_sync_vllm_weights`` after the
# optimizer step on the main thread.
#
# - Async prefetch OFF (synchronous mode): TRL's stock
# ``_generate_single_turn`` calls ``sync_weights`` once per step
# boundary. There's no BG thread to race with, and
# ``_maybe_sync_vllm_weights`` short-circuits with
# ``if not async_prefetch: return``, so we MUST wire the stock
# hook directly to our LoRA sync helper — otherwise nothing ever
# pushes weights to vLLM and the trainer becomes a no-op (vLLM
# keeps serving the base model, every rollout in every group
# produces identical outputs, advantages are zero, optimizer
# step gets skipped, repeat).
if not getattr(self, "_patched_sync_weights", False):
if self.use_vllm and hasattr(self, "vllm_generation"):
if getattr(self.args, "vllm_lora_sync", False):
# No-op: LoRA sync is driven by _maybe_sync_vllm_weights
self.vllm_generation.sync_weights = lambda: None
if getattr(self.args, "async_prefetch", False):
# Async: drive sync from main thread via
# _maybe_sync_vllm_weights instead.
self.vllm_generation.sync_weights = lambda: None
else:
# Sync mode: TRL's _generate_single_turn already
# calls sync_weights once per step boundary. Wire
# it directly to our LoRA filesystem sync helper.
sync_helper = self._sync_lora_adapter
def _lora_filesystem_sync():
sync_helper()
self.vllm_generation.sync_weights = _lora_filesystem_sync
self._patched_sync_weights = True
else:
from accelerate.utils import is_peft_model

View File

@@ -2,17 +2,35 @@
# Copyright (c) Axolotl AI
# Licensed under the Apache License, Version 2.0
from . import layers
from .lora_ops import ParallelExperts
from .parallel_experts import flatten_sort_count, parallel_linear
from .parallel_linear_lora import ScatterMoELoRA, parallel_linear_lora
from .lora_layout import (
peft_down_proj_lora_to_scattermoe,
peft_lora_B_to_scattermoe,
peft_lora_to_scattermoe,
validate_scattermoe_lora_shapes,
)
__all__ = [
"layers",
"ParallelExperts",
"flatten_sort_count",
"parallel_linear",
"ScatterMoELoRA",
"parallel_linear_lora",
"lora_ops",
"peft_down_proj_lora_to_scattermoe",
"peft_lora_B_to_scattermoe",
"peft_lora_to_scattermoe",
"validate_scattermoe_lora_shapes",
]
try:
from . import layers
from .lora_ops import ParallelExperts
from .parallel_experts import flatten_sort_count, parallel_linear
from .parallel_linear_lora import ScatterMoELoRA, parallel_linear_lora
except ModuleNotFoundError as exc:
if exc.name != "triton":
raise
else:
__all__ += [
"layers",
"ParallelExperts",
"flatten_sort_count",
"parallel_linear",
"ScatterMoELoRA",
"parallel_linear_lora",
"lora_ops",
]

View File

@@ -35,81 +35,19 @@ import torch
from torch import nn
from torch.nn import functional as F
from .lora_layout import (
peft_down_proj_lora_to_scattermoe,
peft_lora_B_to_scattermoe,
peft_lora_to_scattermoe,
)
from .parallel_experts import flatten_sort_count, parallel_linear
from .parallel_linear_lora import get_lora_params_from_wrapper, parallel_linear_lora
# =============================================================================
# LoRA layout conversion utilities (peft <-> scattermoe)
# =============================================================================
def peft_lora_B_to_scattermoe(peft_B, num_experts, rank):
"""Convert peft rank-major lora_B ``[out, E*r]`` to scattermoe
expert-major ``[N, r*E]``.
peft reshapes B to ``[out, r, E]`` (rank-major).
scattermoe slices B as ``[:, e*r:(e+1)*r]`` (expert-major).
"""
N = peft_B.shape[0]
return (
peft_B.reshape(N, rank, num_experts)
.permute(0, 2, 1)
.contiguous()
.reshape(N, num_experts * rank)
)
def peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
"""Convert peft LoRA weights to scattermoe layout (with A<->B swap).
peft operates on the parameter in its native storage layout ``[E, dim1, dim2]``
where ``in_features=dim1, out_features=dim2``. ScatterMoE transposes the
parameter (``W = param.transpose(2, 1)``) giving ``[E, dim2, dim1]`` with
``K=dim2, N=dim1``. Because of this transposition, peft's A and B roles
are swapped relative to scattermoe's convention.
peft gives:
lora_A ``[r*E, dim1]``, lora_B ``[dim2, r*E]``
scattermoe needs:
lora_A ``[r*E, K=dim2]``, lora_B ``[N=dim1, r*E]``
This function swaps A<->B and converts B from rank-major to expert-major.
Uses vectorized tensor operations (no Python loop over experts).
Works for **both** gate_up_proj and down_proj since the transposition
issue is the same for any parameter.
"""
peft_B_em = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
dim1 = peft_A.shape[1] # peft in_features -> scattermoe N
dim2 = peft_B_em.shape[0] # peft out_features -> scattermoe K
# smoe_A: per expert, transpose B_e [dim2, r] -> [r, dim2]
# [dim2, E*r] -> [dim2, E, r] -> [E, r, dim2] -> [E*r, dim2]
smoe_A = (
peft_B_em.reshape(dim2, num_experts, rank)
.permute(1, 2, 0)
.contiguous()
.reshape(rank * num_experts, dim2)
)
# smoe_B: per expert, transpose A_e [r, dim1] -> [dim1, r]
# [E*r, dim1] -> [E, r, dim1] -> [dim1, E, r] -> [dim1, E*r]
smoe_B = (
peft_A.reshape(num_experts, rank, dim1)
.permute(2, 0, 1)
.contiguous()
.reshape(dim1, num_experts * rank)
)
return smoe_A, smoe_B
def peft_down_proj_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
"""Deprecated alias for :func:`peft_lora_to_scattermoe`."""
return peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank)
__all__ = [
"peft_down_proj_lora_to_scattermoe",
"peft_lora_B_to_scattermoe",
"peft_lora_to_scattermoe",
]
# =============================================================================
# ParamWrapper unwrapping
@@ -199,7 +137,7 @@ def _unwrap_experts_lora(experts_module):
if gup is not None:
num_experts = gup.shape[0]
# Extract gate_up_proj LoRA (needs A<->B swap due to transposition)
# Extract gate_up_proj LoRA
gup_lora = None
gup_wrapper = wrappers.get("gate_up_proj")
if gup_wrapper is not None:
@@ -208,7 +146,7 @@ def _unwrap_experts_lora(experts_module):
rank = lora_A.shape[0] // num_experts
gup_lora = _convert_smoe_lora(lora_A, lora_B, num_experts, rank, scaling)
# Extract down_proj LoRA (needs A<->B swap due to transposition)
# Extract down_proj LoRA
down_lora = None
down_wrapper = wrappers.get("down_proj")
if down_wrapper is not None:

View File

@@ -0,0 +1,78 @@
# SPDX-License-Identifier: Apache-2.0
# Copyright (c) Axolotl AI
# Licensed under the Apache License, Version 2.0
"""Pure tensor layout helpers for ScatterMoE LoRA weights."""
def peft_lora_B_to_scattermoe(peft_B, num_experts, rank):
"""Convert peft rank-major lora_B ``[out, E*r]`` to scattermoe
expert-major ``[N, r*E]``.
peft reshapes B to ``[out, r, E]`` (rank-major).
scattermoe slices B as ``[:, e*r:(e+1)*r]`` (expert-major).
"""
N = peft_B.shape[0]
return (
peft_B.reshape(N, rank, num_experts)
.permute(0, 2, 1)
.contiguous()
.reshape(N, num_experts * rank)
)
def peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
"""Convert peft LoRA weights to scattermoe layout.
peft operates on the parameter in its native storage layout ``[E, dim1, dim2]``
where ``out_features=dim1, in_features=dim2``. ScatterMoE transposes the
parameter (``W = param.transpose(2, 1)``), giving ``[E, dim2, dim1]`` with
``K=dim2, N=dim1``.
peft gives:
lora_A ``[r*E, dim2]``, lora_B ``[dim1, r*E]``
scattermoe needs:
lora_A ``[r*E, K=dim2]``, lora_B ``[N=dim1, r*E]``
peft's A already matches ScatterMoE's A shape. Only B needs conversion from
peft's rank-major layout to ScatterMoE's expert-major layout.
"""
smoe_A = peft_A
smoe_B = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
return smoe_A, smoe_B
def peft_down_proj_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
"""Deprecated alias for :func:`peft_lora_to_scattermoe`."""
return peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank)
def validate_scattermoe_lora_shapes(expert_weights, lora_A, lora_B):
"""Validate LoRA tensor layout before dispatching ScatterMoE kernels."""
E, K, N = expert_weights.shape
if lora_A.dim() != 2 or lora_B.dim() != 2:
raise ValueError(
"ScatterMoE LoRA expects 2D lora_A and lora_B tensors, got "
f"lora_A={tuple(lora_A.shape)} and lora_B={tuple(lora_B.shape)}."
)
if lora_A.size(0) % E != 0:
raise ValueError(
"ScatterMoE LoRA expects lora_A rows to be divisible by the number "
f"of experts ({E}), got lora_A={tuple(lora_A.shape)}."
)
rank = lora_A.size(0) // E
expected_A = (E * rank, K)
expected_B = (N, E * rank)
if tuple(lora_A.shape) != expected_A or tuple(lora_B.shape) != expected_B:
raise ValueError(
"Invalid ScatterMoE LoRA layout for expert_weights "
f"{tuple(expert_weights.shape)}. Expected lora_A={expected_A} and "
f"lora_B={expected_B}, got lora_A={tuple(lora_A.shape)} and "
f"lora_B={tuple(lora_B.shape)}. For PEFT target_parameters, keep "
"lora_A as [E*r, K] and only convert lora_B from rank-major to "
"expert-major layout."
)

View File

@@ -34,6 +34,7 @@ from .kernels.lora_ops import (
scatter2scatter_lora,
scatter2scatter_lora_dX,
)
from .lora_layout import validate_scattermoe_lora_shapes
class ScatterMoELoRA(torch.autograd.Function):
@@ -422,11 +423,6 @@ def get_lora_params_from_wrapper(module) -> tuple:
return lora_A, lora_B, scaling
# =============================================================================
# Drop-in replacement for parallel_linear
# =============================================================================
def parallel_linear_lora(
inputs: torch.Tensor,
expert_weights: torch.Tensor,
@@ -451,6 +447,7 @@ def parallel_linear_lora(
Otherwise falls back to standard scatter2scatter.
"""
if lora_A is not None and lora_B is not None:
validate_scattermoe_lora_shapes(expert_weights, lora_A, lora_B)
return ScatterMoELoRA.apply(
inputs,
expert_weights,

View File

@@ -110,11 +110,36 @@ class NemoGymDataProducer(GRPODataProducer):
item["agent_ref"] = full_item["agent_ref"]
dataset_items.append(item)
# Expand by num_generations (agent produces one rollout per call)
expanded_items = []
for item in dataset_items:
for _ in range(self._num_generations):
expanded_items.append(item)
# NOTE: do NOT re-expand by num_generations here.
# ``RepeatSampler(mini_repeat_count=num_generations)`` already
# yields ``num_generations`` consecutive copies of each unique
# prompt, so ``inputs`` is a list of ``(unique_prompts_per_rank *
# num_generations)`` items — one entry per rollout. Expanding
# again here would fire ``num_generations^2`` rollouts per
# prompt per rank and make every step dogpile on a handful of
# tasks.
expanded_items = dataset_items
# Diagnostic: log what this rank is about to fire.
try:
import collections
iid_counts: collections.Counter[str | None] = collections.Counter()
for it in dataset_items:
iid_counts[
(it.get("responses_create_params", {}).get("metadata") or {}).get(
"instance_id"
)
] += 1
LOG.info(
"[RANK:%d] produce(): firing %d agent /run calls covering %d unique prompts: %s",
trainer.accelerator.process_index,
len(dataset_items),
len(iid_counts),
list(iid_counts.most_common(5)),
)
except Exception:
pass
# Call NeMo Gym agents
loop = asyncio.new_event_loop()
@@ -140,6 +165,7 @@ class NemoGymDataProducer(GRPODataProducer):
logprobs_list = []
rewards_list = []
num_turns_list: list[int] = []
for resp in responses:
parsed = _parse_agent_response(resp, eos_token_id)
prompt_ids_list.append(parsed["prompt_ids"])
@@ -147,6 +173,7 @@ class NemoGymDataProducer(GRPODataProducer):
env_mask_list.append(parsed["env_mask"])
logprobs_list.append(parsed["logprobs"])
rewards_list.append(parsed["reward"])
num_turns_list.append(parsed.get("num_turns", 0))
# Pad to tensors
prompt_ids = [torch.tensor(ids, device=device) for ids in prompt_ids_list]
@@ -179,22 +206,48 @@ class NemoGymDataProducer(GRPODataProducer):
tool_mask = [torch.tensor(m, device=device) for m in env_mask_list]
tool_mask = pad(tool_mask, padding_value=1, padding_side="right")
# Inject rewards into inputs so _compute_deferred_scores can use them
# The deferred scoring path calls _calculate_rewards which reads reward_funcs.
# Our passthrough reward_fn reads "env_reward" from kwargs.
# Inject per-rollout reward + num_turns into each input. Since
# ``RepeatSampler`` already yields ``num_generations`` copies of
# each prompt, ``inputs`` has ONE entry per rollout (matching
# ``rewards_list`` 1:1). No per-prompt grouping happens here —
# GRPO advantage normalization is the trainer's job downstream.
assert len(inputs) == len(rewards_list), (
f"rewards/inputs length mismatch: "
f"{len(rewards_list)} rewards vs {len(inputs)} inputs"
)
for i, inp in enumerate(inputs):
# Each input gets rewards for its num_generations rollouts
start = i * self._num_generations
end = start + self._num_generations
inp["env_reward"] = rewards_list[start:end]
inp["env_reward"] = rewards_list[i]
inp["num_turns"] = num_turns_list[i]
# Expand inputs to match expanded rollouts (num_generations copies)
expanded_inputs = []
for inp in inputs:
for g in range(self._num_generations):
expanded_inp = dict(inp)
expanded_inp["env_reward"] = inp["env_reward"][g]
expanded_inputs.append(expanded_inp)
# One expanded_input per rollout (already correct count because
# inputs has num_generations copies baked in by the sampler).
expanded_inputs = [dict(inp) for inp in inputs]
# Log rollout-level stats to wandb from rank 0. These are the
# true agent-side metrics (not the tokenized TRL view) — so
# num_turns reflects how many /run iterations each rollout
# actually took before finishing or hitting max_turns.
if is_main and num_turns_list:
try:
import wandb
if wandb.run is not None:
import statistics as _stats
nonzero = sum(1 for r in rewards_list if r > 0)
log_payload = {
"rollout/num_turns/mean": float(_stats.mean(num_turns_list)),
"rollout/num_turns/min": float(min(num_turns_list)),
"rollout/num_turns/max": float(max(num_turns_list)),
"rollout/reward/mean": float(_stats.mean(rewards_list)),
"rollout/reward/nonzero_frac": (
nonzero / len(rewards_list) if rewards_list else 0.0
),
"rollout/n_samples": float(len(rewards_list)),
}
wandb.log(log_payload, commit=False)
except Exception as exc: # never let metric logging break training
LOG.warning("rollout wandb log failed: %s", exc)
# Decode completions for reward functions
completions = trainer.processing_class.batch_decode(

View File

@@ -19,6 +19,7 @@ Supports two modes:
from __future__ import annotations
import os
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Union
from axolotl.integrations.base import BasePlugin
@@ -30,6 +31,107 @@ if TYPE_CHECKING:
LOG = get_logger(__name__)
# ---- vLLM weight-sync transport probe ------------------------------------
@dataclass
class VLLMWeightSyncCapabilities:
"""What weight-sync routes a vLLM server actually exposes.
Discovered once at ``pre_model_load`` time by fetching the server's
``/openapi.json``. Drives the transport-selection table below.
"""
nccl: bool = False # /init_communicator/ + /update_named_param/
lora_filesystem: bool = False # /v1/load_lora_adapter (vLLM native)
lora_axolotl: bool = False # /set_lora_adapter/ (axolotl serve_lora extension)
http_full: bool = False # /http_update_weights/ (axolotl serve_lora extension)
probed: bool = False
probe_error: str | None = None
routes: list[str] = field(default_factory=list)
@property
def any_full_param_sync(self) -> bool:
"""True if at least one transport can push full-model weights."""
return self.nccl or self.http_full
@property
def any_lora_sync(self) -> bool:
"""True if at least one transport can push LoRA adapters."""
return self.lora_filesystem or self.lora_axolotl or self.nccl
def probe_vllm_weight_sync(
base_url: str, timeout: float = 5.0
) -> VLLMWeightSyncCapabilities:
"""Detect which weight-sync routes the configured vLLM server exposes.
Uses the server's FastAPI ``/openapi.json`` — every weight-sync transport
we care about is mounted as a POST route there. Falls back to all-False
on any error so the caller can still decide what to do (typically: raise
a clear error rather than silently no-op).
"""
import requests
caps = VLLMWeightSyncCapabilities()
try:
r = requests.get(f"{base_url.rstrip('/')}/openapi.json", timeout=timeout)
r.raise_for_status()
spec = r.json()
routes = sorted((spec.get("paths") or {}).keys())
caps.routes = routes
caps.nccl = "/init_communicator/" in routes and "/update_named_param/" in routes
caps.lora_filesystem = "/v1/load_lora_adapter" in routes
caps.lora_axolotl = "/set_lora_adapter/" in routes
caps.http_full = "/http_update_weights/" in routes
caps.probed = True
except Exception as exc:
caps.probe_error = f"{type(exc).__name__}: {exc}"
LOG.warning(
"NeMo Gym: failed to probe vLLM /openapi.json at %s%s. "
"Will fall back to LoRA-only behavior.",
base_url,
caps.probe_error,
)
return caps
def select_weight_sync_transport(
caps: VLLMWeightSyncCapabilities,
*,
has_lora: bool,
vllm_lora_sync_pref: bool,
) -> str:
"""Pick the right transport for a (server caps, model type) combo.
Returns one of: ``"lora_filesystem"``, ``"nccl"``, ``"http_full"``, or
``"none"``. The caller decides what to do with ``"none"`` (typically:
raise an error explaining the misconfiguration).
Selection table:
LoRA model + lora endpoint + lora-sync pref → lora_filesystem
LoRA model + lora endpoint → lora_filesystem
LoRA model + nccl endpoint → nccl (broadcast merged adapter)
Full model + nccl endpoint → nccl
Full model + http endpoint → http_full
anything else → none
"""
if has_lora:
if (caps.lora_filesystem or caps.lora_axolotl) and vllm_lora_sync_pref:
return "lora_filesystem"
if caps.lora_filesystem or caps.lora_axolotl:
return "lora_filesystem"
if caps.nccl:
return "nccl"
return "none"
# Full-parameter model
if caps.nccl:
return "nccl"
if caps.http_full:
return "http_full"
return "none"
class NemoGymPlugin(BasePlugin):
"""Plugin for NVIDIA NeMo Gym integration with Axolotl.
@@ -50,37 +152,69 @@ class NemoGymPlugin(BasePlugin):
self._reward_fn = None
self._dataset_lookup = None
self._agent_servers = {}
self._vllm_caps: VLLMWeightSyncCapabilities | None = None
def get_input_args(self):
return "axolotl.integrations.nemo_gym.NemoGymArgs"
def pre_model_load(self, cfg):
"""Apply monkeypatches before trainer creation."""
"""Probe vLLM weight-sync routes and conditionally bypass NCCL init.
Replaces the previous unconditional ``init_communicator`` monkey-patch
with a probe of the configured vLLM server's ``/openapi.json``. We only
bypass NCCL init when the server we're talking to actually lacks the
``/init_communicator/`` route (i.e. stock ``vllm serve``); against
TRL/axolotl serve modules that DO expose NCCL routes, we leave the
standard TRL flow alone so full-finetune training can sync weights.
"""
if not cfg.nemo_gym_enabled:
return
# Always skip NCCL communicator init in NeMo Gym mode.
# NeMo Gym uses its own vLLM server (standard OpenAI API), not the TRL
# colocate/NCCL path. The NCCL init fails with vLLM V1 and standard servers.
trl_cfg = getattr(cfg, "trl", None)
if trl_cfg and getattr(trl_cfg, "vllm_mode", "server") == "server":
if not (trl_cfg and getattr(trl_cfg, "vllm_mode", "server") == "server"):
return
host = getattr(trl_cfg, "vllm_server_host", None) or "127.0.0.1"
port = getattr(trl_cfg, "vllm_server_port", None) or 8000
base_url = f"http://{host}:{port}"
self._vllm_caps = probe_vllm_weight_sync(base_url)
if self._vllm_caps.probed:
LOG.info(
"NeMo Gym: vLLM weight-sync probe @ %s — nccl=%s lora_native=%s "
"lora_axolotl=%s http_full=%s",
base_url,
self._vllm_caps.nccl,
self._vllm_caps.lora_filesystem,
self._vllm_caps.lora_axolotl,
self._vllm_caps.http_full,
)
# Only bypass NCCL init when the server doesn't speak it. If NCCL is
# available we leave VLLMClient.init_communicator alone so the
# standard TRL sync flow can run for full-parameter training.
if not self._vllm_caps.nccl:
self._patch_skip_nccl_init()
def _patch_skip_nccl_init(self):
"""Monkeypatch VLLMClient.init_communicator to no-op.
NeMo Gym uses its own vLLM server (standard OpenAI API or custom LoRA
serve script). The NCCL communicator is not needed and fails with both
vLLM V1 engine and standard OpenAI server mode.
Only called when the configured vLLM server doesn't expose
``/init_communicator/`` (e.g. stock ``vllm serve``). In that case
TRL's standard ``init_communicator`` would 404 inside trainer
construction; we no-op it so the LoRA filesystem path can install
its own sync in ``post_trainer_create``.
"""
try:
from trl.generation.vllm_client import VLLMClient
VLLMClient._original_init_communicator = VLLMClient.init_communicator
VLLMClient.init_communicator = lambda self, **kwargs: LOG.info(
"Skipping NCCL init_communicator (LoRA sync mode)"
"Skipping NCCL init_communicator (server has no /init_communicator/)"
)
LOG.info(
"Patched VLLMClient.init_communicator to no-op (server has no NCCL routes)"
)
LOG.info("Patched VLLMClient.init_communicator to no-op for LoRA sync")
except Exception as exc:
LOG.warning(f"Failed to patch VLLMClient: {exc}")
@@ -234,30 +368,80 @@ class NemoGymPlugin(BasePlugin):
verify_timeout = cfg.nemo_gym_verify_timeout or 30
multi_turn = cfg.nemo_gym_multi_turn or False
# Handle weight sync. NeMo Gym skips NCCL init, so we need to either:
# - Install LoRA sync (when vllm_lora_sync=True)
# - Or no-op sync_weights (when using standard vLLM server)
# Pick a weight-sync transport based on what the configured vLLM
# server actually exposes (see ``pre_model_load`` probe) and what
# kind of model we're training. The selection table is documented
# in ``select_weight_sync_transport``.
trl_cfg = getattr(cfg, "trl", None)
if hasattr(trainer, "vllm_generation") and trainer.vllm_generation:
vllm_gen = trainer.vllm_generation
if trl_cfg and getattr(trl_cfg, "vllm_lora_sync", False):
adapter = getattr(cfg, "adapter", None)
has_lora = adapter in ("lora", "qlora")
vllm_lora_sync_pref = bool(
trl_cfg and getattr(trl_cfg, "vllm_lora_sync", False)
)
caps = self._vllm_caps or VLLMWeightSyncCapabilities()
transport = select_weight_sync_transport(
caps,
has_lora=has_lora,
vllm_lora_sync_pref=vllm_lora_sync_pref,
)
if transport == "lora_filesystem":
self._setup_lora_sync(trainer)
# Verify the vLLM server supports runtime LoRA loading
self._check_lora_endpoint(vllm_gen)
else:
# No NCCL, no LoRA sync — skip all weight sync paths
vllm_gen.sync_weights = lambda: LOG.debug(
"Weight sync skipped (NeMo Gym mode)"
LOG.info("NeMo Gym weight sync: LoRA filesystem")
elif transport == "nccl":
# Standard TRL NCCL path. We leave ``VLLMClient.init_communicator``
# alone (pre_model_load only patched it when the probe found no
# NCCL route) so the trainer's normal weight-sync flow runs.
LOG.info(
"NeMo Gym weight sync: NCCL (server exposes /init_communicator/)"
)
type(vllm_gen).sync_weights = lambda self: LOG.debug(
"Weight sync skipped (NeMo Gym mode)"
elif transport == "http_full":
# Full-parameter HTTP sync — implementation lands in step 3.
# For now, fail loudly so users know the path is detected but
# not yet wired up, instead of silently no-oping like before.
raise NotImplementedError(
"NeMo Gym + full fine-tune + HTTP weight sync is detected "
"but the client-side sync helper is not yet implemented "
"(planned). Use `adapter: lora|qlora` for now, or use a "
"vLLM serve module that exposes /init_communicator/ for "
"NCCL sync."
)
# Also patch the async trainer's internal sync method
if hasattr(trainer, "_maybe_sync_vllm_weights"):
trainer._maybe_sync_vllm_weights = lambda: LOG.debug(
"Async weight sync skipped (NeMo Gym mode)"
else: # transport == "none"
# No viable sync path. Build a precise error so the user knows
# exactly what's missing and how to fix it.
if not caps.probed:
msg = (
"could not probe the vLLM server's "
f"/openapi.json: {caps.probe_error}. "
"Verify that vLLM is reachable at "
f"{getattr(trl_cfg, 'vllm_server_host', '?')}:"
f"{getattr(trl_cfg, 'vllm_server_port', '?')}."
)
LOG.info("Disabled weight sync (NeMo Gym mode, no LoRA sync)")
elif has_lora:
msg = (
"the vLLM server has neither NCCL routes "
"(/init_communicator/) nor a LoRA-loading route "
"(/v1/load_lora_adapter or /set_lora_adapter/). "
"Restart vLLM with `--enable-lora --max-lora-rank N "
"VLLM_ALLOW_RUNTIME_LORA_UPDATING=1` for the stock "
"server, or use `axolotl vllm-serve` for the "
"NCCL-capable serve module."
)
else:
msg = (
"the vLLM server exposes no full-parameter sync route "
"(/init_communicator/ for NCCL or /http_update_weights/ "
"for HTTP). Use `axolotl vllm-serve` (which has both) "
"or set `adapter: lora|qlora`."
)
raise ValueError(
f"NeMo Gym: no usable weight-sync transport — {msg} Without "
"weight sync the trainer's gradient updates never reach the "
"rollout policy (functionally a no-op trainer)."
)
if multi_turn:
self._wire_multi_turn(cfg, trainer, model_name, verify_timeout)

View File

@@ -130,21 +130,41 @@ def start_servers(
)
def get_server_configs(head_port: int = 11000) -> dict:
def get_server_configs(head_port: int = 11000, timeout: float = 30.0) -> dict:
"""Fetch the global config from the NeMo Gym head server.
Retries up to 3 times with exponential backoff. The default per-attempt
timeout is 30s (raised from the original 5s) because head servers can
be slow to respond when they're concurrently serving rollouts from a
prior training run. A 5s timeout was empirically too tight to survive
a kill-and-relaunch cycle.
Returns:
Dict mapping server_name -> server config.
"""
response = requests.get(
f"http://127.0.0.1:{head_port}/global_config_dict_yaml", timeout=5
url = f"http://127.0.0.1:{head_port}/global_config_dict_yaml"
last_exc: Exception | None = None
for attempt in (1, 2, 3):
try:
response = requests.get(url, timeout=timeout)
response.raise_for_status()
result = yaml.safe_load(response.text)
# NeMo Gym head server double-encodes: YAML string inside a YAML string
if isinstance(result, str):
result = yaml.safe_load(result)
return result
except (requests.exceptions.RequestException, OSError) as exc:
last_exc = exc
LOG.warning(
"NeMo Gym head probe attempt %d/3 failed: %s. Retrying...",
attempt,
type(exc).__name__,
)
if attempt < 3:
time.sleep(2.0 * attempt)
raise RuntimeError(
f"NeMo Gym head server at {url} did not respond after 3 attempts: {last_exc}"
)
response.raise_for_status()
result = yaml.safe_load(response.text)
# NeMo Gym head server double-encodes: YAML string inside a YAML string
if isinstance(result, str):
result = yaml.safe_load(result)
return result
def get_agent_servers(

View File

@@ -53,6 +53,7 @@ def _rms_norm_rope_forward_kernel(
RSTD_ptr,
RSTD_row_stride,
n_cols,
n_rot,
n_heads,
eps,
HAS_WEIGHT: tl.constexpr,
@@ -60,28 +61,35 @@ def _rms_norm_rope_forward_kernel(
):
"""
Fused forward:
x_norm = x / rms(x) [* weight] (RMSNorm)
y = x_norm * cos + rotate_half(x_norm) * sin (RoPE)
x_norm = x / rms(x) [* weight] (RMSNorm, full n_cols)
y[..., :n_rot] = rope(x_norm[..., :n_rot])
y[..., n_rot:] = x_norm[..., n_rot:] (pass-through for partial rotary)
rotate_half swaps first/second halves and negates the first:
rotate_half([a, b]) = [-b, a]
rotate_half swaps first/second halves and negates the first, restricted
to the rotary span [0, n_rot):
rotate_half([a, b]) = [-b, a] where len(a) = len(b) = n_rot/2
For the partial-rotary pass-through region we load cos with default 1.0
and sin with default 0.0 outside [0, n_rot), so the same formula
`Y = X_norm * cos + X_rot_norm * sin` collapses to `Y = X_norm`.
cos/sin are indexed by row_idx // n_heads to handle per-head broadcast
(cos/sin have shape (B*S, D) while X has shape (B*S*H, D)).
(cos/sin have shape (B*S, n_rot) while X has shape (B*S*H, n_cols)).
"""
row_idx = tl.program_id(0).to(tl.int64)
# cos/sin row: divide by n_heads since cos/sin are (B*S, D)
# cos/sin row: divide by n_heads since cos/sin are (B*S, n_rot)
cs_row_idx = row_idx // n_heads
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < n_cols
half_dim = n_cols // 2
rot_mask_col = col_offsets < n_rot
half_rot = n_rot // 2
# Load input row
X_row = tl.load(X_ptr + row_idx * X_row_stride + col_offsets, mask=mask, other=0)
X_dtype = X_row.dtype
X_fp32 = X_row.to(tl.float32)
# RMSNorm: compute 1/rms
# RMSNorm: compute 1/rms over the full row (rotary + pass-through)
mean_sq = tl.sum(X_fp32 * X_fp32, axis=0) / n_cols
rstd = rsqrt(mean_sq + eps)
tl.store(RSTD_ptr + row_idx * RSTD_row_stride, rstd)
@@ -94,33 +102,38 @@ def _rms_norm_rope_forward_kernel(
W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0).to(tl.float32)
X_norm = X_norm * W_row
# RoPE: load cos/sin (broadcast across heads)
# RoPE: load cos/sin (broadcast across heads). For col >= n_rot we get
# cos=1, sin=0 so the formula leaves X_norm untouched.
cos_row = tl.load(
COS_ptr + cs_row_idx * COS_row_stride + col_offsets, mask=mask, other=0
COS_ptr + cs_row_idx * COS_row_stride + col_offsets,
mask=rot_mask_col,
other=1.0,
).to(tl.float32)
sin_row = tl.load(
SIN_ptr + cs_row_idx * SIN_row_stride + col_offsets, mask=mask, other=0
SIN_ptr + cs_row_idx * SIN_row_stride + col_offsets,
mask=rot_mask_col,
other=0.0,
).to(tl.float32)
# rotate_half: for col < half_dim, take -X_norm[col + half_dim]
# for col >= half_dim, take X_norm[col - half_dim]
# rotate_half within [0, n_rot):
# for col < half_rot: take -X_norm[col + half_rot]
# for col in [half_rot, n_rot): take X_norm[col - half_rot]
# For col >= n_rot the rotation is irrelevant (sin = 0 zeros it out).
rot_offsets = tl.where(
col_offsets < half_dim, col_offsets + half_dim, col_offsets - half_dim
col_offsets < half_rot, col_offsets + half_rot, col_offsets - half_rot
)
rot_mask = rot_offsets < n_cols
rot_load_mask = (rot_offsets < n_cols) & rot_mask_col
X_rot = tl.load(
X_ptr + row_idx * X_row_stride + rot_offsets, mask=rot_mask & mask, other=0
X_ptr + row_idx * X_row_stride + rot_offsets, mask=rot_load_mask, other=0
).to(tl.float32)
# Re-normalize the rotated values
X_rot_norm = X_rot * rstd
if HAS_WEIGHT:
W_rot = tl.load(W_ptr + rot_offsets, mask=rot_mask & mask, other=0).to(
tl.float32
)
W_rot = tl.load(W_ptr + rot_offsets, mask=rot_load_mask, other=0).to(tl.float32)
X_rot_norm = X_rot_norm * W_rot
# Negate the first half (rotate_half negates x2, which becomes the first half)
sign = tl.where(col_offsets < half_dim, -1.0, 1.0)
sign = tl.where(col_offsets < half_rot, -1.0, 1.0)
X_rot_norm = X_rot_norm * sign
# Final RoPE: y = x_norm * cos + rotate_half(x_norm) * sin
@@ -153,13 +166,21 @@ def _rms_norm_rope_backward_kernel(
dW_row_stride,
n_rows,
n_cols,
n_rot,
n_heads,
rows_per_program,
HAS_WEIGHT: tl.constexpr,
BLOCK_SIZE: tl.constexpr,
):
"""
Backward for Y = RoPE(RMSNorm(X, W))
Backward for Y = RoPE(RMSNorm(X, W)) with optional partial rotary
(`n_rot <= n_cols`).
For col < n_rot the standard RoPE adjoint applies. For col >= n_rot the
output is just the normalized row, so dN[col] = dY[col] (achieved by
loading cos with default 1.0 and forcing the rotate-half contribution
to zero outside the rotary span).
cos/sin indexed by row_idx // n_heads for per-head broadcast.
"""
row_block_id = tl.program_id(0).to(tl.int64)
@@ -167,7 +188,8 @@ def _rms_norm_rope_backward_kernel(
row_end = min((row_block_id + 1) * rows_per_program, n_rows)
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < n_cols
half_dim = n_cols // 2
rot_mask_col = col_offsets < n_rot
half_rot = n_rot // 2
dW_acc = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
@@ -186,33 +208,37 @@ def _rms_norm_rope_backward_kernel(
rstd = tl.load(RSTD_ptr + row_idx * RSTD_row_stride)
cos_row = tl.load(
COS_ptr + cs_row_idx * COS_row_stride + col_offsets, mask=mask, other=0
COS_ptr + cs_row_idx * COS_row_stride + col_offsets,
mask=rot_mask_col,
other=1.0,
).to(tl.float32)
# dN = dY * cos + rotate_half^T(dY * sin)
# dN = dY * cos + rotate_half^T(dY * sin) (within the rotary span)
# rotate_half^T([a, b]) = [b, -a] (adjoint of rotate_half)
#
# Compute rotate_half_transpose(dY * sin) by loading dY and sin at
# rotated offsets directly: dY[rot] * sin[rot] * adj_sign
# This is equivalent to rotating (dY * sin) because the rotation
# just permutes which elements are multiplied.
# For col >= n_rot the formula must collapse to dN = dY (since the
# forward is just a pass-through). cos defaults to 1.0 above; the
# rotate-half contribution is masked to zero below.
rot_offsets = tl.where(
col_offsets < half_dim, col_offsets + half_dim, col_offsets - half_dim
col_offsets < half_rot, col_offsets + half_rot, col_offsets - half_rot
)
rot_mask = rot_offsets < n_cols
rot_load_mask = (rot_offsets < n_cols) & rot_mask_col
dY_rot = tl.load(
dY_ptr + row_idx * dY_row_stride + rot_offsets,
mask=rot_mask & mask,
mask=rot_load_mask,
other=0,
).to(tl.float32)
sin_rot = tl.load(
SIN_ptr + cs_row_idx * SIN_row_stride + rot_offsets,
mask=rot_mask & mask,
mask=rot_load_mask,
other=0,
).to(tl.float32)
adj_sign = tl.where(col_offsets < half_dim, 1.0, -1.0)
dN = dY_row * cos_row + dY_rot * sin_rot * adj_sign
adj_sign = tl.where(col_offsets < half_rot, 1.0, -1.0)
rotate_term = dY_rot * sin_rot * adj_sign
# Zero out rotate-half contribution outside the rotary span.
rotate_term = tl.where(rot_mask_col, rotate_term, 0.0)
dN = dY_row * cos_row + rotate_term
# Pre-weight normalized: n = rstd * x
n = X_row * rstd
@@ -241,15 +267,17 @@ def _rms_norm_rope_backward_kernel(
)
def rms_norm_rope_forward(X, W, cos, sin, eps, n_heads):
def rms_norm_rope_forward(X, W, cos, sin, eps, n_heads, n_rot):
"""
Args:
X: (B*S*H, head_dim) — contiguous, flattened from (B, S, H, D)
W: (head_dim,) or None — RMSNorm weight
cos: (B*S, head_dim) — position embeddings (broadcast across heads)
sin: (B*S, head_dim) — position embeddings (broadcast across heads)
cos: (B*S, n_rot) — position embeddings (broadcast across heads)
sin: (B*S, n_rot) — position embeddings (broadcast across heads)
eps: float
n_heads: int — number of attention heads (for cos/sin indexing)
n_rot: int — rotary dim (== head_dim for full rotary, < head_dim for
partial rotary). Must be even and ``<= head_dim``.
Returns:
Y, X_saved, RSTD, BLOCK_SIZE, num_warps
"""
@@ -273,6 +301,7 @@ def rms_norm_rope_forward(X, W, cos, sin, eps, n_heads):
RSTD,
RSTD.stride(0),
n_cols,
n_rot,
n_heads,
eps,
HAS_WEIGHT=has_weight,
@@ -282,7 +311,9 @@ def rms_norm_rope_forward(X, W, cos, sin, eps, n_heads):
return Y, X, RSTD, BLOCK_SIZE, num_warps
def rms_norm_rope_backward(dY, X, W, cos, sin, RSTD, n_heads, BLOCK_SIZE, num_warps):
def rms_norm_rope_backward(
dY, X, W, cos, sin, RSTD, n_heads, n_rot, BLOCK_SIZE, num_warps
):
n_rows, n_cols = dY.shape
has_weight = W is not None
@@ -315,6 +346,7 @@ def rms_norm_rope_backward(dY, X, W, cos, sin, RSTD, n_heads, BLOCK_SIZE, num_wa
_dW.stride(0),
n_rows,
n_cols,
n_rot,
n_heads,
rows_per_program,
HAS_WEIGHT=has_weight,
@@ -329,13 +361,14 @@ def rms_norm_rope_backward(dY, X, W, cos, sin, RSTD, n_heads, BLOCK_SIZE, num_wa
class FusedRMSNormRoPEFunction(torch.autograd.Function):
@staticmethod
@ensure_contiguous
def forward(ctx, X, W, cos, sin, eps, n_heads):
def forward(ctx, X, W, cos, sin, eps, n_heads, n_rot):
"""
X: (B*S*H, head_dim)
W: (head_dim,) or None
cos: (B*S, head_dim) — broadcast across heads
sin: (B*S, head_dim) — broadcast across heads
X: (B*S*H, head_dim)
W: (head_dim,) or None
cos: (B*S, n_rot) — broadcast across heads
sin: (B*S, n_rot) — broadcast across heads
n_heads: int
n_rot: int — rotary dim (<= head_dim)
"""
Y, X_saved, RSTD, BLOCK_SIZE, num_warps = rms_norm_rope_forward(
X,
@@ -344,11 +377,13 @@ class FusedRMSNormRoPEFunction(torch.autograd.Function):
sin,
eps,
n_heads,
n_rot,
)
ctx.eps = eps
ctx.BLOCK_SIZE = BLOCK_SIZE
ctx.num_warps = num_warps
ctx.n_heads = n_heads
ctx.n_rot = n_rot
ctx.has_weight = W is not None
ctx.save_for_backward(X_saved, W, cos, sin, RSTD)
return Y
@@ -365,21 +400,26 @@ class FusedRMSNormRoPEFunction(torch.autograd.Function):
sin,
RSTD,
ctx.n_heads,
ctx.n_rot,
ctx.BLOCK_SIZE,
ctx.num_warps,
)
return dX, dW, None, None, None, None
return dX, dW, None, None, None, None, None
def fused_rms_norm_rope(x, weight, cos, sin, eps=1e-6):
"""
Apply fused RMSNorm + RoPE.
Apply fused RMSNorm + (partial) RoPE.
Args:
x: (batch, seq_len, num_heads, head_dim) — after projection + view
weight: (head_dim,) — RMSNorm weight, or None for no-scale norm
cos: (batch, seq_len, head_dim) — from RotaryEmbedding
sin: (batch, seq_len, head_dim) — from RotaryEmbedding
cos: (batch, seq_len, n_rot) — from RotaryEmbedding. ``n_rot``
must be even and ``<= head_dim``. When ``n_rot < head_dim``
the trailing ``head_dim - n_rot`` columns are RMSNorm-only
(partial-rotary pass-through), matching stock Gemma 4 with
``partial_rotary_factor < 1.0``.
sin: (batch, seq_len, n_rot) — same shape as ``cos``
eps: float — RMSNorm epsilon
Returns:
@@ -387,14 +427,38 @@ def fused_rms_norm_rope(x, weight, cos, sin, eps=1e-6):
"""
shape = x.shape # (B, S, H, D)
B, S, H, D = shape
n_rot = cos.shape[-1]
if sin.shape[-1] != n_rot:
raise ValueError(
f"cos and sin must have the same last dim, got cos={cos.shape[-1]} "
f"sin={sin.shape[-1]}"
)
if n_rot > D:
raise ValueError(f"rotary dim ({n_rot}) cannot exceed head_dim ({D})")
if n_rot % 2 != 0:
raise ValueError(f"rotary dim must be even, got {n_rot}")
# Flatten to 2D: (B*S*H, D)
x_flat = x.reshape(-1, D).contiguous()
# Flatten cos/sin to (B*S, D) — the kernel will handle per-head broadcast
# by dividing the row_idx by H to get the cos/sin row
cos_flat = cos.reshape(B * S, D).contiguous()
sin_flat = sin.reshape(B * S, D).contiguous()
# cos/sin may broadcast over the batch dim (e.g. (1, S, n_rot) when
# all sequences share the same rotary positions). The kernel needs a
# dense (B*S, n_rot) buffer so that row_idx // n_heads maps cleanly
# onto a single (b, s) pair, so expand-then-contiguous to materialize
# the per-batch broadcast. Expand is a no-op when B == cos.shape[0].
if cos.shape[0] != B:
if cos.shape[0] != 1:
raise ValueError(
f"cos/sin batch dim ({cos.shape[0]}) must be 1 or equal "
f"to x batch dim ({B})"
)
cos = cos.expand(B, S, n_rot)
sin = sin.expand(B, S, n_rot)
cos_flat = cos.reshape(B * S, n_rot).contiguous()
sin_flat = sin.reshape(B * S, n_rot).contiguous()
y_flat = FusedRMSNormRoPEFunction.apply(x_flat, weight, cos_flat, sin_flat, eps, H)
y_flat = FusedRMSNormRoPEFunction.apply(
x_flat, weight, cos_flat, sin_flat, eps, H, n_rot
)
return y_flat.view(shape)

View File

@@ -156,6 +156,14 @@ class PatchManager:
# which would clobber any earlier fix.
self._fix_nemotron_h_conversion_mapping()
# Gemma 4 hybrid attention runs here in post-build (NOT post-load):
# the per-layer ``self_attn.config._attn_implementation="sdpa"``
# override needs to walk the raw model tree, which is broken by
# the post-load PEFT wrapping. The accompanying
# ``patch_gemma4_hybrid_mask`` monkey-patch is module-level and
# installation-time-independent, so both halves of the fix live
# cleanly in the same call even though one is instance-scoped
# and the other is module-scoped.
self._apply_gemma_hybrid_attention(model)
self._finalize_moe_expert_quantization(model)
@@ -173,12 +181,23 @@ class PatchManager:
which exceeds flash attention's supported size. This patch loads the model
with flash_attention_2 for the sliding window layers (head_dim=256), then
gives each global layer a shallow-copied config with _attn_implementation="sdpa".
We also install :func:`axolotl.monkeypatch.gemma4_hybrid_mask.patch_gemma4_hybrid_mask`
which fixes the corresponding mask construction inside
``Gemma4TextModel.forward``. Without it, the per-layer SDPA config
override is not enough — the forward still builds a 2D FA2-format mask
at the model level and the SDPA layers crash at long context lengths
with ``RuntimeError: The expanded size of the tensor ... must match``.
"""
if not self.cfg.gemma4_hybrid_attn_impl:
return
import copy
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
patch_gemma4_hybrid_mask()
# Navigate to the module that has 'layers' - varies by model structure:
# Gemma4ForConditionalGeneration -> .model (Gemma4Model) -> .language_model (Gemma4TextModel) -> .layers
# Gemma4ForCausalLM -> .model (Gemma4TextModel) -> .layers
@@ -392,6 +411,14 @@ class PatchManager:
patch_qwen3_5_vlm_flash_attention()
if self.cfg.model_config_type in ("gemma4", "gemma4_text"):
# The fused attn path is now compatible with
# ``gemma4_hybrid_attn_impl``: the kernel handles partial
# rotary (cos.shape[-1] < head_dim) and the fused forward
# mirrors the current ``Gemma4TextAttention.forward`` API
# for shared kv (read from / write to
# ``past_key_values.shared_layers``). See
# ``src/axolotl/kernels/GEMMA4_FUSED_ROPE_HYBRID_ATTN_BUG.md``
# for the history.
from axolotl.monkeypatch.models.gemma4.fused_attn import (
patch_gemma4_fused_attn,
)

View File

@@ -0,0 +1,115 @@
"""Hybrid attention mask fix for Gemma 4.
Gemma 4 has full-attention (global) layers with ``head_dim=512`` which
exceeds flash-attention-2's supported size. Axolotl's hybrid-attention
patch in ``patch_manager._apply_gemma_hybrid_attention`` works around
this by forcing ``_attn_implementation="sdpa"`` on each global layer's
``self_attn.config``, leaving sliding-window layers on FA2.
The per-layer config override alone is insufficient, however:
``Gemma4TextModel.forward`` builds a single ``causal_mask_mapping`` dict
using the **model-level** config and passes the mapped mask to each
decoder layer. With FA2 still set at the model level, the ``full_attention``
entry in that mapping is a 2D mask (FA2 format), but SDPA needs a 4D mask.
The global layers then fail with::
RuntimeError: The expanded size of the tensor (S) must match the existing
size (B) at non-singleton dimension 2. Target sizes: [B, H, S, S]. Tensor
sizes: [B, S]
...when the sequence length grows past roughly 7k tokens.
This module fixes the symptom by monkey-patching ``create_causal_mask`` in
``transformers.models.gemma4.modeling_gemma4``'s module namespace — NOT
the original in ``masking_utils``. The wrapper forces
``_attn_implementation="sdpa"`` on a shallow-copied config before calling
through, so the ``full_attention`` mask built inside ``Gemma4TextModel.forward``
is always 4D/SDPA-compatible. ``create_sliding_window_causal_mask`` is left
alone, so sliding-window layers continue to receive FA2-format masks.
The patch is idempotent. Install once per process, before any Gemma 4
forward pass runs.
"""
from __future__ import annotations
import copy
from typing import Any
from axolotl.utils.logging import get_logger
LOG = get_logger(__name__)
_PATCH_APPLIED = False
def patch_gemma4_hybrid_mask() -> bool:
"""Install the Gemma 4 hybrid-attention mask fix.
Returns ``True`` if the patch was installed (or was already installed),
``False`` if the target module could not be imported (e.g. transformers
version predates Gemma 4) — in which case nothing is done and the
caller can continue unaffected.
"""
global _PATCH_APPLIED
if _PATCH_APPLIED:
return True
try:
from transformers.models.gemma4 import modeling_gemma4
except ImportError:
LOG.debug(
"gemma4_hybrid_mask: transformers.models.gemma4 not importable, "
"skipping. This is fine for non-Gemma4 training."
)
return False
if not hasattr(modeling_gemma4, "create_causal_mask"):
LOG.warning(
"gemma4_hybrid_mask: modeling_gemma4 has no 'create_causal_mask' "
"binding, skipping. Transformers API may have changed."
)
return False
original = modeling_gemma4.create_causal_mask
def hybrid_create_causal_mask(config: Any, *args: Any, **kwargs: Any):
"""Wrapper that forces SDPA format for the full-attention mask.
The global layers were patched to SDPA by
``_apply_gemma_hybrid_attention``, so their mask must be 4D. The
original ``create_causal_mask`` dispatches on
``config._attn_implementation``; we shadow that with a local
override.
"""
sdpa_config = copy.copy(config)
sdpa_config._attn_implementation = "sdpa"
return original(sdpa_config, *args, **kwargs)
# Preserve the original reference on the wrapper for tests / teardown.
hybrid_create_causal_mask._axolotl_original = original # type: ignore[attr-defined]
modeling_gemma4.create_causal_mask = hybrid_create_causal_mask
_PATCH_APPLIED = True
LOG.info(
"gemma4_hybrid_mask: patched modeling_gemma4.create_causal_mask to "
"force SDPA-format masks for full-attention layers"
)
return True
def unpatch_gemma4_hybrid_mask() -> None:
"""Restore the original ``create_causal_mask``. Useful for tests."""
global _PATCH_APPLIED
if not _PATCH_APPLIED:
return
try:
from transformers.models.gemma4 import modeling_gemma4
except ImportError:
_PATCH_APPLIED = False
return
current = modeling_gemma4.create_causal_mask
original = getattr(current, "_axolotl_original", None)
if original is not None:
modeling_gemma4.create_causal_mask = original
_PATCH_APPLIED = False

View File

@@ -24,7 +24,15 @@ def patch_tiled_mlp(model_type, use_original_mlp=True, cfg_num_shards=None):
module_path = f"transformers.models.{model_type}.modeling_{model_type}"
model_cls_prefix, _ = get_causal_lm_model_cls_prefix(model_type)
module = __import__(module_path, fromlist=[f"{model_cls_prefix}MLP"])
mlp_cls = getattr(module, f"{model_cls_prefix}MLP")
# Some multimodal wrappers (e.g. Gemma 4) name the MLP class
# ``{prefix}TextMLP`` rather than ``{prefix}MLP`` because the
# language-side module is separated from the vision tower. Try
# both names before giving up.
mlp_cls = getattr(
module,
f"{model_cls_prefix}MLP",
None,
) or getattr(module, f"{model_cls_prefix}TextMLP")
if use_original_mlp:
mlp_forward = mlp_cls.forward

View File

@@ -320,6 +320,15 @@ def main(script_args: ScriptArguments):
# --- Active LoRA state (shared across endpoints via closure) ---
active_lora: dict = {"request": None}
# Serializes access to the worker pipe. The underlying
# multiprocessing.Connection is a single full-duplex stream shared
# across all HTTP handlers; concurrent requests interleave bytes on
# the wire and corrupt the pickle framing (seen as
# ``UnpicklingError: pickle data was truncated``). Any endpoint that
# does ``conn.send(...); conn.recv()`` MUST hold this lock across
# the round-trip so only one inflight call at a time per pipe.
worker_pipe_lock = asyncio.Lock()
# ------------------------------------------------------------------
# LoRA-specific endpoints
# ------------------------------------------------------------------
@@ -631,6 +640,150 @@ def main(script_args: ScriptArguments):
},
}
@app.post("/v1/completions")
async def openai_completions(request_body: dict):
"""OpenAI-compatible text-completions endpoint.
Accepts either a string ``prompt`` or a list-of-int
``prompt_token_ids`` (as the text-completions spec allows). Routes
to the internal vLLM generate method with the active LoRA adapter
and returns an OpenAI /v1/completions-shaped response including
per-choice ``prompt_token_ids``, ``generation_token_ids``, and
``generation_log_probs`` for NeMo Gym agents that need raw
tokens + logprobs.
"""
import uuid
prompt_raw = request_body.get("prompt")
temperature = request_body.get("temperature", 1.0)
max_tokens = request_body.get("max_tokens", 512)
top_p = request_body.get("top_p", 1.0)
n = request_body.get("n", 1)
logprobs = request_body.get("logprobs") or 0
stop_token_ids = request_body.get("stop_token_ids") or None
# Accept either a string or a list[int] token id prompt. Lists
# must contain ints only (raise on lists of strings so callers get
# a clear error). Also accept [[int, int, ...]] nesting for the
# rare case callers pass a single-prompt batch.
if (
isinstance(prompt_raw, list)
and prompt_raw
and isinstance(prompt_raw[0], list)
):
prompt_raw = prompt_raw[0]
prompt_dict: dict[str, Any] = {}
if isinstance(prompt_raw, list):
prompt_dict = {"prompt_token_ids": prompt_raw}
elif isinstance(prompt_raw, str):
prompt_dict = {"prompt": prompt_raw}
else:
return {
"error": {
"message": ("prompt must be a string or a list of token ids"),
"type": "invalid_request",
}
}
generation_kwargs: dict[str, Any] = {
"n": n,
"temperature": temperature,
"top_p": top_p,
"max_tokens": max_tokens,
"logprobs": logprobs,
}
if stop_token_ids:
generation_kwargs["stop_token_ids"] = stop_token_ids
sampling_params = SamplingParams(
**{k: v for k, v in generation_kwargs.items() if v is not None}
)
chunked = chunk_list([prompt_dict], script_args.data_parallel_size)
# Hold the pipe lock across send+recv — concurrent requests would
# otherwise interleave pickle frames on the worker connection.
async with worker_pipe_lock:
for conn, chunk in zip(connections, chunked, strict=True):
if not chunk:
chunk = [{"prompt": "<placeholder>"}]
kwargs = {
"prompts": chunk,
"sampling_params": sampling_params,
"lora_request": active_lora["request"],
}
conn.send({"type": "call", "method": "generate", "kwargs": kwargs})
loop = asyncio.get_running_loop()
all_outputs = await asyncio.gather(
*(loop.run_in_executor(None, safe_recv, conn) for conn in connections)
)
all_outputs = [o for o, c in zip(all_outputs, chunked, strict=True) if c]
for o in all_outputs:
if isinstance(o, dict) and "error" in o:
raise RuntimeError(f"vLLM worker error: {o['error']}")
all_outputs = list(chain.from_iterable(all_outputs))
if not all_outputs:
return {"choices": [], "model": script_args.model}
choices = []
for i, output in enumerate(all_outputs):
for j, out in enumerate(output.outputs):
text = out.text
# OpenAI-style `logprobs` block for text-completions:
# { "tokens": [...], "token_logprobs": [...] }
lp_block = None
if out.logprobs:
tokens_str: list[str] = []
token_lps: list[float] = []
for step in out.logprobs:
chosen = next(iter(step.values()))
tokens_str.append(getattr(chosen, "decoded_token", "") or "")
token_lps.append(float(chosen.logprob))
lp_block = {
"tokens": tokens_str,
"token_logprobs": token_lps,
}
choice = {
"index": i * n + j,
"text": text,
"finish_reason": "stop"
if out.finish_reason == "stop"
else "length",
"logprobs": lp_block,
# NeMo-Gym / retrace agent extras — preserved on the
# choice so callers with raw-token pipelines don't
# have to re-tokenize.
"prompt_token_ids": output.prompt_token_ids,
"generation_token_ids": list(out.token_ids),
"generation_log_probs": (
[float(next(iter(lp.values())).logprob) for lp in out.logprobs]
if out.logprobs
else []
),
}
choices.append(choice)
prompt_tokens = len(all_outputs[0].prompt_token_ids) if all_outputs else 0
completion_tokens = sum(
len(out.token_ids) for o in all_outputs for out in o.outputs
)
return {
"id": f"cmpl-{uuid.uuid4().hex[:8]}",
"object": "text_completion",
"model": script_args.model,
"choices": choices,
"usage": {
"prompt_tokens": prompt_tokens,
"completion_tokens": completion_tokens,
"total_tokens": prompt_tokens + completion_tokens,
},
}
# --- Weight sync endpoints (legacy fallback, same as TRL) ---
@app.post("/init_communicator/")

View File

@@ -770,6 +770,88 @@ class RLValidationMixin:
)
return data
@model_validator(mode="before")
@classmethod
def check_grpo_batch_size_divisibility(cls, data):
"""Surface GRPO batch-shape mismatches at config-parse time.
TRL's GRPOTrainer requires that the per-step generation batch size be
evenly divisible by ``num_generations`` so that every prompt can be
replicated exactly ``num_generations`` times. The runtime check inside
``GRPOTrainer.__init__`` only fires after the model has been loaded —
too late and too cryptic for the user. We replicate the check here so
the failure is immediate and actionable.
Also enforces:
- ``num_generations >= 2`` (group-relative advantage needs variance)
- ``effective_gbs >= num_generations * world_size`` when capabilities
indicate multiple ranks (each rank needs at least one full group)
"""
if data.get("rl") != "grpo":
return data
trl_cfg = data.get("trl") or {}
num_gen = trl_cfg.get("num_generations")
if num_gen is None:
# TRL's own default is 8 — but if the user didn't set it, we
# don't have enough info to validate anything. Let TRL's own
# init handle the default-vs-batch interaction.
return data
if num_gen < 2:
raise ValueError(
f"GRPO requires `trl.num_generations >= 2` (got {num_gen}). "
"With num_generations=1, every group has zero advantage and "
"the policy never updates."
)
explicit_gbs = trl_cfg.get("generation_batch_size")
if explicit_gbs is not None:
effective_gbs = int(explicit_gbs)
gbs_source = "trl.generation_batch_size"
else:
mb = data.get("micro_batch_size") or 1
ga = data.get("gradient_accumulation_steps") or 1
effective_gbs = int(mb) * int(ga)
gbs_source = f"micro_batch_size ({mb}) * gradient_accumulation_steps ({ga})"
if effective_gbs % num_gen != 0:
# Suggest the smallest GA bump that fixes it for the common case
# where the user hasn't set generation_batch_size explicitly.
hint = ""
if explicit_gbs is None:
from math import gcd
mb_val = int(data.get("micro_batch_size") or 1)
# smallest GA such that mb*GA is a multiple of num_gen
lcm = num_gen * mb_val // gcd(num_gen, mb_val)
suggested_ga = lcm // mb_val
hint = (
f" Smallest fix: set `gradient_accumulation_steps: "
f"{suggested_ga}` (so micro_batch_size * GA = "
f"{mb_val * suggested_ga} is a multiple of {num_gen})."
)
raise ValueError(
f"GRPO: generation batch size must be divisible by "
f"`trl.num_generations`. Got effective_gbs={effective_gbs} "
f"(from {gbs_source}) and num_generations={num_gen}.{hint}"
)
# Multi-rank check: each rank must receive at least one full group
# per step. Without `capabilities` populated yet (mode='before'), we
# fall back to user-set distributed fields.
world_size = (
(data.get("capabilities") or {}).get("n_gpu") or data.get("world_size") or 1
)
if world_size and world_size > 1 and effective_gbs < num_gen * world_size:
raise ValueError(
f"GRPO with world_size={world_size} requires effective_gbs "
f">= num_generations * world_size = {num_gen * world_size}, "
f"got {effective_gbs}. Increase gradient_accumulation_steps "
f"or micro_batch_size."
)
return data
class OptimizationValidationMixin:
"""Validation methods related to optimization and performance."""

View File

@@ -216,5 +216,197 @@ class TestValidateQuantPatchRestore(unittest.TestCase):
self.assertIs(_trainer_module.validate_quantization_for_training, original)
class TestVllmLoraSyncPatch(unittest.TestCase):
"""The ``_generate_single_turn`` patch wires sync_weights to the right place.
These tests exercise the patch-installation branch in isolation. They build
a stub trainer with just enough attributes to look like
``AsyncGRPOTrainer`` for the duration of the relevant code path.
Background — there are two correct behaviors and we historically had a bug
where both modes used the same one:
- Async prefetch ON: the BG generation thread can't safely call
sync_weights mid-rollout. We no-op the stock hook and drive sync from
the main thread via ``_maybe_sync_vllm_weights``.
- Async prefetch OFF: TRL's stock ``_generate_single_turn`` already
calls ``sync_weights`` once per step boundary on the main thread. We
wire that hook directly to ``_sync_lora_adapter`` because
``_maybe_sync_vllm_weights`` short-circuits when async is off.
Before the fix, both modes installed ``lambda: None``, so sync mode never
pushed any LoRA adapter to vLLM and the trainer was a no-op.
"""
@staticmethod
def _make_stub_trainer(*, vllm_lora_sync, async_prefetch):
from axolotl.core.trainers.grpo.async_trainer import (
AsyncGRPOTrainer,
)
class FakeArgs:
pass
args = FakeArgs()
args.vllm_lora_sync = vllm_lora_sync
args.async_prefetch = async_prefetch
class FakeVllmGen:
sync_weights = staticmethod(lambda: None)
model = MagicMock()
# Use object.__new__ so we don't run __init__ (which needs a real
# model, dataset, etc.). We only need the `_generate_single_turn`
# method's patch branch to run, so we set up the minimum state.
trainer = object.__new__(AsyncGRPOTrainer)
trainer.args = args
trainer.use_vllm = True
trainer.vllm_generation = FakeVllmGen()
trainer._patched_sync_weights = False
# Spy on _sync_lora_adapter so we can assert it's the function the
# hook delegates to in sync mode.
trainer._sync_lora_adapter = MagicMock(name="_sync_lora_adapter_spy")
trainer._sync_peft_weights_no_merge = MagicMock(
name="_sync_peft_weights_no_merge_spy"
)
return trainer
@staticmethod
def _run_patch_branch(trainer):
"""Execute just the sync_weights-patching branch in isolation.
We can't easily call the real ``_generate_single_turn`` because it
does a full vLLM generate. Instead we copy the exact branch out of
the source so the test verifies the same logic the trainer runs.
"""
if not getattr(trainer, "_patched_sync_weights", False):
if trainer.use_vllm and hasattr(trainer, "vllm_generation"):
if getattr(trainer.args, "vllm_lora_sync", False):
if getattr(trainer.args, "async_prefetch", False):
trainer.vllm_generation.sync_weights = lambda: None
else:
sync_helper = trainer._sync_lora_adapter
def _lora_filesystem_sync():
sync_helper()
trainer.vllm_generation.sync_weights = _lora_filesystem_sync
trainer._patched_sync_weights = True
def test_sync_mode_with_lora_sync_wires_to_sync_lora_adapter(self):
trainer = self._make_stub_trainer(vllm_lora_sync=True, async_prefetch=False)
self._run_patch_branch(trainer)
assert trainer._patched_sync_weights is True
# Trigger the patched hook — it must call _sync_lora_adapter.
trainer.vllm_generation.sync_weights()
trainer._sync_lora_adapter.assert_called_once()
def test_async_mode_with_lora_sync_installs_noop_hook(self):
trainer = self._make_stub_trainer(vllm_lora_sync=True, async_prefetch=True)
self._run_patch_branch(trainer)
assert trainer._patched_sync_weights is True
# Hook must be a no-op so BG-thread generation doesn't fight the
# main-thread optimizer step over the model weights.
trainer.vllm_generation.sync_weights()
trainer._sync_lora_adapter.assert_not_called()
def test_sync_mode_with_lora_sync_does_not_call_during_install(self):
"""Installing the patch should not pre-emptively sync."""
trainer = self._make_stub_trainer(vllm_lora_sync=True, async_prefetch=False)
self._run_patch_branch(trainer)
# _sync_lora_adapter should only be called when the patched hook
# itself is invoked (e.g., from TRL's _generate_single_turn).
trainer._sync_lora_adapter.assert_not_called()
def test_patch_is_idempotent(self):
trainer = self._make_stub_trainer(vllm_lora_sync=True, async_prefetch=False)
self._run_patch_branch(trainer)
first_hook = trainer.vllm_generation.sync_weights
# Second call must not re-patch (otherwise we'd lose the original).
self._run_patch_branch(trainer)
assert trainer.vllm_generation.sync_weights is first_hook
class TestMaybeSyncVllmWeightsIntervalDefault(unittest.TestCase):
"""``_maybe_sync_vllm_weights`` must not crash when interval is unset.
Before the fix, ``step % self.args.vllm_sync_interval`` would TypeError
on the very first call when ``vllm_sync_interval`` was ``None`` (which
is the default for any config that doesn't explicitly set it). We now
fall back to interval=1 so unset means "sync every step", matching the
behavior of TRL's own ``_generate_single_turn``.
"""
@staticmethod
def _make_stub_trainer(interval, async_prefetch):
from axolotl.core.trainers.grpo.async_trainer import (
AsyncGRPOTrainer,
)
class FakeArgs:
pass
args = FakeArgs()
args.async_prefetch = async_prefetch
args.vllm_sync_interval = interval
args.vllm_lora_sync = True
class FakeState:
global_step = 1
trainer = object.__new__(AsyncGRPOTrainer)
trainer.args = args
trainer.use_vllm = True
trainer.state = FakeState()
trainer._last_synced_step = 0
trainer._sync_lora_adapter = MagicMock(name="sync_spy")
return trainer
def test_interval_none_in_async_mode_does_not_crash(self):
trainer = self._make_stub_trainer(interval=None, async_prefetch=True)
from axolotl.core.trainers.grpo.async_trainer import (
AsyncGRPOTrainer,
)
# Should not raise TypeError — defaults to every-step sync
AsyncGRPOTrainer._maybe_sync_vllm_weights(trainer)
trainer._sync_lora_adapter.assert_called_once()
def test_sync_mode_drives_sync(self):
"""Sync mode must fire ``_sync_lora_adapter`` from ``_maybe_sync_vllm_weights``.
The previous behavior (early return when ``not async_prefetch``)
assumed TRL's stock ``_generate_single_turn`` would handle sync.
That's true for vanilla GRPO but FALSE for NeMo Gym multi-turn
where the data producer bypasses ``_generate_single_turn``
entirely. Without this trigger no sync ever happens and the
trainer becomes a no-op.
"""
trainer = self._make_stub_trainer(interval=1, async_prefetch=False)
from axolotl.core.trainers.grpo.async_trainer import (
AsyncGRPOTrainer,
)
AsyncGRPOTrainer._maybe_sync_vllm_weights(trainer)
trainer._sync_lora_adapter.assert_called_once()
def test_async_mode_with_explicit_interval_respects_modulo(self):
trainer = self._make_stub_trainer(interval=4, async_prefetch=True)
from axolotl.core.trainers.grpo.async_trainer import (
AsyncGRPOTrainer,
)
# global_step=1, interval=4 → 1 % 4 != 0 → no sync
AsyncGRPOTrainer._maybe_sync_vllm_weights(trainer)
trainer._sync_lora_adapter.assert_not_called()
# global_step=4 → 4 % 4 == 0 → sync
trainer.state.global_step = 4
AsyncGRPOTrainer._maybe_sync_vllm_weights(trainer)
trainer._sync_lora_adapter.assert_called_once()
if __name__ == "__main__":
unittest.main()

View File

@@ -54,25 +54,7 @@ except (ImportError, ModuleNotFoundError):
)
def peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
peft_B_em = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
K_inter, N_hidden = peft_B.shape[0], peft_A.shape[1]
smoe_A = torch.zeros(
rank * num_experts,
K_inter,
device=peft_A.device,
dtype=peft_A.dtype,
)
smoe_B = torch.zeros(
N_hidden,
rank * num_experts,
device=peft_A.device,
dtype=peft_A.dtype,
)
for e in range(num_experts):
s = e * rank
smoe_A[s : s + rank, :] = peft_B_em[:, s : s + rank].T
smoe_B[:, s : s + rank] = peft_A[s : s + rank, :].T
return smoe_A, smoe_B
return peft_A, peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
def _unwrap_experts_lora(experts_module):
return experts_module, None, None
@@ -145,11 +127,7 @@ def scattermoe_lora_B_to_peft(smoe_B, num_experts, rank):
def peft_gate_up_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
"""Convert peft LoRA for gate_up_proj to scattermoe layout.
Both gate_up_proj and down_proj need the A<->B swap because
scattermoe transposes the parameter (W = param.T).
"""
"""Convert peft LoRA for gate_up_proj to scattermoe layout."""
return peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank)
@@ -322,14 +300,16 @@ class TestLoRABLayoutConversion:
hidden, inter = 32, 16
scaling = 2.0
peft_A = torch.randn(E * r, hidden)
peft_B = torch.randn(inter, E * r)
peft_A = torch.randn(E * r, inter)
peft_B = torch.randn(hidden, E * r)
A_r = peft_A.reshape(E, r, hidden)
B_r = peft_B.reshape(inter, r, E)
delta_peft = torch.einsum("o r e, e r i -> e i o", B_r, A_r) * scaling
A_r = peft_A.reshape(E, r, inter)
B_r = peft_B.reshape(hidden, r, E)
delta_peft = torch.einsum("o r e, e r i -> e o i", B_r, A_r) * scaling
smoe_A, smoe_B = peft_lora_to_scattermoe(peft_A, peft_B, E, r)
assert smoe_A.shape == (E * r, inter)
assert smoe_B.shape == (hidden, E * r)
for e in range(E):
A_e = smoe_A[e * r : (e + 1) * r, :]
B_e = smoe_B[:, e * r : (e + 1) * r]
@@ -342,27 +322,26 @@ class TestLoRABLayoutConversion:
"""Verify gate_up_proj LoRA conversion with non-square dims (Qwen3-like).
gate_up_proj param: [E, 2*inter, hidden].
peft: in_features=2*inter, out_features=hidden.
peft lora_A: [r*E, 2*inter], lora_B: [hidden, r*E].
peft: in_features=hidden, out_features=2*inter.
peft lora_A: [r*E, hidden], lora_B: [2*inter, r*E].
scattermoe W = param.T = [E, hidden, 2*inter], K=hidden, N=2*inter.
scattermoe needs: lora_A [r*E, K=hidden], lora_B [N=2*inter, r*E].
Uses non-square dims (hidden=32 != 2*inter=24) to catch A<->B swap bugs.
Uses non-square dims (hidden=32 != 2*inter=24) to catch layout bugs.
"""
E, r = 4, 2
hidden, inter = 32, 12 # 2*inter=24 != hidden=32
scaling = 2.0
# peft assigns: in_features=2*inter, out_features=hidden
peft_A = torch.randn(E * r, 2 * inter) # [r*E, in_features=2*inter]
peft_B = torch.randn(hidden, E * r) # [out_features=hidden, r*E]
# peft assigns: in_features=hidden, out_features=2*inter
peft_A = torch.randn(E * r, hidden) # [r*E, in_features=hidden]
peft_B = torch.randn(2 * inter, E * r) # [out_features=2*inter, r*E]
# peft delta via einsum: "o r e, e r i -> e i o"
A_r = peft_A.reshape(E, r, 2 * inter)
B_r = peft_B.reshape(hidden, r, E)
delta_peft = torch.einsum("o r e, e r i -> e i o", B_r, A_r) * scaling
# delta_peft[e] has shape [in_features, out_features] = [2*inter, hidden]
A_r = peft_A.reshape(E, r, hidden)
B_r = peft_B.reshape(2 * inter, r, E)
delta_peft = torch.einsum("o r e, e r i -> e o i", B_r, A_r) * scaling
# delta_peft[e] has shape [out_features, in_features] = [2*inter, hidden]
# = param[e] shape [2*inter, hidden]
smoe_A, smoe_B = peft_gate_up_lora_to_scattermoe(peft_A, peft_B, E, r)
@@ -422,22 +401,22 @@ class TestPeftLoRAWeightExtraction:
)
# gate_up_proj [E, 2*inter, hidden]
# peft: in_features=2*inter (dim 1), out_features=hidden (dim 2)
# peft: in_features=hidden (last dim), out_features=2*inter (middle dim)
assert trainable[
"base_model.model.moe.experts.base_layer.lora_A.default.weight"
].shape == (E * r, 2 * config.intermediate_size)
assert trainable[
"base_model.model.moe.experts.base_layer.lora_B.default.weight"
].shape == (config.hidden_size, E * r)
# down_proj [E, hidden, inter]
# peft: in_features=hidden (dim 1), out_features=inter (dim 2)
assert trainable[
"base_model.model.moe.experts.lora_A.default.weight"
].shape == (E * r, config.hidden_size)
assert trainable[
"base_model.model.moe.experts.base_layer.lora_B.default.weight"
].shape == (2 * config.intermediate_size, E * r)
# down_proj [E, hidden, inter]
# peft: in_features=inter (last dim), out_features=hidden (middle dim)
assert trainable[
"base_model.model.moe.experts.lora_A.default.weight"
].shape == (E * r, config.intermediate_size)
assert trainable[
"base_model.model.moe.experts.lora_B.default.weight"
].shape == (config.intermediate_size, E * r)
].shape == (config.hidden_size, E * r)
@requires_cuda
def test_peft_forward_runs(self):
@@ -488,27 +467,29 @@ class TestPeftLoRAWeightExtraction:
assert gup_lora is not None, "gate_up_proj LoRA not detected"
assert down_lora is not None, "down_proj LoRA not detected"
# Check shapes (after peft->scattermoe conversion with A<->B swap)
# gate_up_proj W = param.T = [E, hidden, 2*inter], K=hidden, N=2*inter
# Check shapes after peft->scattermoe conversion.
# gate_up_proj: peft A [E*r, hidden] / B [2*inter, E*r]
# scattermoe: smoe_A [E*r, hidden], smoe_B [2*inter, E*r]
E, r = config.num_experts, 4
gup_A, gup_B, gup_s = gup_lora
assert gup_A.shape == (E * r, config.hidden_size), (
f"gate_up_proj smoe_A: expected [r*E, K=hidden]={(E * r, config.hidden_size)}, "
f"gate_up_proj smoe_A: expected [r*E, hidden]={(E * r, config.hidden_size)}, "
f"got {gup_A.shape}"
)
assert gup_B.shape == (2 * config.intermediate_size, E * r), (
f"gate_up_proj smoe_B: expected [N=2*inter, r*E]="
f"gate_up_proj smoe_B: expected [2*inter, r*E]="
f"{(2 * config.intermediate_size, E * r)}, got {gup_B.shape}"
)
# down_proj W = param.T = [E, inter, hidden], K=inter, N=hidden
# down_proj: peft A [E*r, inter] / B [hidden, E*r]
# scattermoe: smoe_A [E*r, inter], smoe_B [hidden, E*r]
down_A, down_B, down_s = down_lora
assert down_A.shape == (E * r, config.intermediate_size), (
f"down_proj smoe_A: expected [r*E, K=inter]={(E * r, config.intermediate_size)}, "
f"down_proj smoe_A: expected [r*E, inter]={(E * r, config.intermediate_size)}, "
f"got {down_A.shape}"
)
assert down_B.shape == (config.hidden_size, E * r), (
f"down_proj smoe_B: expected [N=hidden, r*E]={(config.hidden_size, E * r)}, "
f"down_proj smoe_B: expected [hidden, r*E]={(config.hidden_size, E * r)}, "
f"got {down_B.shape}"
)

View File

@@ -361,6 +361,329 @@ class TestPluginDefaults(unittest.TestCase):
assert cfg.dataloader_num_workers == 0
class TestSelectWeightSyncTransport(unittest.TestCase):
"""Pure-logic table tests for ``select_weight_sync_transport``."""
def _caps(self, **kwargs):
from axolotl.integrations.nemo_gym.plugin import VLLMWeightSyncCapabilities
c = VLLMWeightSyncCapabilities(probed=True)
for k, v in kwargs.items():
setattr(c, k, v)
return c
def test_lora_with_native_endpoint(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps(lora_filesystem=True)
assert (
select_weight_sync_transport(caps, has_lora=True, vllm_lora_sync_pref=True)
== "lora_filesystem"
)
def test_lora_with_axolotl_endpoint(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps(lora_axolotl=True)
assert (
select_weight_sync_transport(caps, has_lora=True, vllm_lora_sync_pref=False)
== "lora_filesystem"
)
def test_lora_falls_back_to_nccl_when_no_lora_endpoint(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps(nccl=True)
assert (
select_weight_sync_transport(caps, has_lora=True, vllm_lora_sync_pref=False)
== "nccl"
)
def test_full_param_prefers_nccl(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps(nccl=True, http_full=True)
assert (
select_weight_sync_transport(
caps, has_lora=False, vllm_lora_sync_pref=False
)
== "nccl"
)
def test_full_param_falls_back_to_http(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps(http_full=True)
assert (
select_weight_sync_transport(
caps, has_lora=False, vllm_lora_sync_pref=False
)
== "http_full"
)
def test_full_param_no_routes_returns_none(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps() # all False
assert (
select_weight_sync_transport(
caps, has_lora=False, vllm_lora_sync_pref=False
)
== "none"
)
def test_lora_no_routes_returns_none(self):
from axolotl.integrations.nemo_gym.plugin import select_weight_sync_transport
caps = self._caps()
assert (
select_weight_sync_transport(caps, has_lora=True, vllm_lora_sync_pref=True)
== "none"
)
class TestProbeVllmWeightSync(unittest.TestCase):
"""``probe_vllm_weight_sync`` reads a vLLM ``/openapi.json`` and reports caps."""
def test_stock_vllm_with_lora_enabled(self):
"""Stock ``vllm serve --enable-lora`` exposes only LoRA endpoints."""
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import probe_vllm_weight_sync
spec = {
"paths": {
"/v1/models": {"get": {}},
"/v1/load_lora_adapter": {"post": {}},
"/v1/unload_lora_adapter": {"post": {}},
"/v1/completions": {"post": {}},
}
}
with patch("requests.get") as mock_get:
mock_get.return_value.raise_for_status = lambda: None
mock_get.return_value.json = lambda: spec
caps = probe_vllm_weight_sync("http://localhost:8000")
assert caps.probed is True
assert caps.lora_filesystem is True
assert caps.lora_axolotl is False
assert caps.nccl is False
assert caps.http_full is False
def test_axolotl_serve_lora_full_capabilities(self):
"""``axolotl vllm-serve`` exposes NCCL + LoRA + HTTP full sync."""
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import probe_vllm_weight_sync
spec = {
"paths": {
"/init_communicator/": {"post": {}},
"/update_named_param/": {"post": {}},
"/batch_update_named_params/": {"post": {}},
"/set_lora_adapter/": {"post": {}},
"/clear_lora_adapter/": {"post": {}},
"/http_update_weights/": {"post": {}},
"/v1/load_lora_adapter": {"post": {}},
}
}
with patch("requests.get") as mock_get:
mock_get.return_value.raise_for_status = lambda: None
mock_get.return_value.json = lambda: spec
caps = probe_vllm_weight_sync("http://localhost:8000")
assert caps.probed is True
assert caps.nccl is True
assert caps.lora_axolotl is True
assert caps.lora_filesystem is True
assert caps.http_full is True
def test_trl_vllm_serve_nccl_only(self):
"""``trl vllm-serve`` exposes NCCL routes but not LoRA filesystem."""
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import probe_vllm_weight_sync
spec = {
"paths": {
"/init_communicator/": {"post": {}},
"/update_named_param/": {"post": {}},
"/batch_update_named_params/": {"post": {}},
"/close_communicator/": {"post": {}},
"/generate/": {"post": {}},
}
}
with patch("requests.get") as mock_get:
mock_get.return_value.raise_for_status = lambda: None
mock_get.return_value.json = lambda: spec
caps = probe_vllm_weight_sync("http://localhost:8000")
assert caps.probed is True
assert caps.nccl is True
assert caps.lora_filesystem is False
assert caps.lora_axolotl is False
assert caps.http_full is False
def test_unreachable_server_records_error(self):
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import probe_vllm_weight_sync
with patch("requests.get") as mock_get:
mock_get.side_effect = ConnectionError("Connection refused")
caps = probe_vllm_weight_sync("http://localhost:9999")
assert caps.probed is False
assert caps.probe_error is not None
assert "ConnectionError" in caps.probe_error
assert caps.nccl is False
assert caps.lora_filesystem is False
class TestPluginWeightSyncEnforcement(unittest.TestCase):
"""End-to-end test of post_trainer_create's transport-selection branch.
The plugin used to silently no-op weight sync when ``vllm_lora_sync: false``,
leaving the trainer learning in isolation while vLLM kept serving the
unmodified base model. After the fix:
- LoRA + LoRA-loading endpoint → installs filesystem LoRA sync
- LoRA + only NCCL endpoint → uses NCCL broadcast
- Full FT + NCCL endpoint → uses NCCL broadcast (standard TRL flow)
- Full FT + HTTP endpoint → raises NotImplementedError (step 3)
- No usable transport → raises ValueError with a precise diagnosis
"""
@staticmethod
def _fake_cfg(adapter, vllm_lora_sync):
class FakeTRL:
pass
class FakeCfg:
pass
trl = FakeTRL()
trl.vllm_lora_sync = vllm_lora_sync
trl.vllm_server_host = "127.0.0.1"
trl.vllm_server_port = 8000
cfg = FakeCfg()
cfg.nemo_gym_enabled = True
cfg.nemo_gym_model_name = None
cfg.base_model = "test/model"
cfg.nemo_gym_verify_timeout = 30
cfg.nemo_gym_multi_turn = True
cfg.adapter = adapter
cfg.trl = trl
return cfg
@staticmethod
def _fake_trainer():
class FakeVLLMGen:
sync_weights = staticmethod(lambda: None)
class FakeTrainer:
vllm_generation = FakeVLLMGen()
return FakeTrainer()
@staticmethod
def _caps(**kwargs):
from axolotl.integrations.nemo_gym.plugin import VLLMWeightSyncCapabilities
c = VLLMWeightSyncCapabilities(probed=True)
for k, v in kwargs.items():
setattr(c, k, v)
return c
def test_lora_with_lora_endpoint_installs_filesystem_sync(self):
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
plugin._vllm_caps = self._caps(lora_filesystem=True)
cfg = self._fake_cfg(adapter="lora", vllm_lora_sync=True)
trainer = self._fake_trainer()
with (
patch.object(plugin, "_setup_lora_sync") as setup,
patch.object(plugin, "_check_lora_endpoint") as check,
patch.object(plugin, "_wire_multi_turn") as wire,
):
plugin.post_trainer_create(cfg, trainer)
setup.assert_called_once()
check.assert_called_once()
wire.assert_called_once()
def test_lora_with_no_routes_raises_with_lora_specific_message(self):
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
plugin._vllm_caps = self._caps() # all False, but probed
cfg = self._fake_cfg(adapter="lora", vllm_lora_sync=False)
trainer = self._fake_trainer()
with self.assertRaises(ValueError) as ctx:
plugin.post_trainer_create(cfg, trainer)
msg = str(ctx.exception)
assert "no-op trainer" in msg
assert "load_lora_adapter" in msg
assert "VLLM_ALLOW_RUNTIME_LORA_UPDATING" in msg
def test_full_finetune_with_nccl_endpoint_uses_nccl(self):
from unittest.mock import patch
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
plugin._vllm_caps = self._caps(nccl=True)
cfg = self._fake_cfg(adapter=None, vllm_lora_sync=False)
trainer = self._fake_trainer()
with patch.object(plugin, "_wire_multi_turn") as wire:
plugin.post_trainer_create(cfg, trainer)
wire.assert_called_once()
def test_full_finetune_with_http_endpoint_not_implemented_yet(self):
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
plugin._vllm_caps = self._caps(http_full=True)
cfg = self._fake_cfg(adapter=None, vllm_lora_sync=False)
trainer = self._fake_trainer()
with self.assertRaises(NotImplementedError) as ctx:
plugin.post_trainer_create(cfg, trainer)
assert "HTTP weight sync" in str(ctx.exception)
def test_full_finetune_with_no_routes_raises_with_full_param_message(self):
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
plugin._vllm_caps = self._caps()
cfg = self._fake_cfg(adapter=None, vllm_lora_sync=False)
trainer = self._fake_trainer()
with self.assertRaises(ValueError) as ctx:
plugin.post_trainer_create(cfg, trainer)
msg = str(ctx.exception)
assert "no-op trainer" in msg
assert "init_communicator" in msg
assert "http_update_weights" in msg
def test_unprobed_caps_raises_with_probe_failure_message(self):
from axolotl.integrations.nemo_gym.plugin import NemoGymPlugin
plugin = NemoGymPlugin()
# Plugin._vllm_caps left as default-None: the post_trainer_create
# branch falls back to a fresh VLLMWeightSyncCapabilities() with
# probed=False, so the error path should mention probing.
cfg = self._fake_cfg(adapter="lora", vllm_lora_sync=True)
trainer = self._fake_trainer()
with self.assertRaises(ValueError) as ctx:
plugin.post_trainer_create(cfg, trainer)
assert "could not probe" in str(ctx.exception)
class TestNemoGymE2E(unittest.TestCase):
"""End-to-end test: data producer → agent (mocked) → parse → tensors → rewards.
@@ -452,19 +775,15 @@ class TestNemoGymE2E(unittest.TestCase):
trainer = self._make_mock_trainer()
producer._trainer = trainer
# Mock the prompt iterator (returns a batch of 1 input)
producer._prompt_iter = iter(
[
[
{
"prompt": [{"role": "user", "content": "Play Wordle!"}],
}
]
]
)
producer._prompt_dl = [
[{"prompt": [{"role": "user", "content": "Play Wordle!"}]}]
# Mock the prompt iterator. RepeatSampler(mini_repeat_count=num_generations)
# pre-expands prompts, so the iterator yields num_generations=2 consecutive
# copies of each unique prompt — one entry per rollout.
_prompt_batch = [
{"prompt": [{"role": "user", "content": "Play Wordle!"}]},
{"prompt": [{"role": "user", "content": "Play Wordle!"}]},
]
producer._prompt_iter = iter([_prompt_batch])
producer._prompt_dl = [_prompt_batch]
# Call produce
result = producer.produce(model=MagicMock(), global_step=1)
@@ -530,10 +849,13 @@ class TestNemoGymE2E(unittest.TestCase):
producer._request_timeout = 30
producer._num_generations = 2
producer._trainer = self._make_mock_trainer()
producer._prompt_iter = iter(
[[{"prompt": [{"role": "user", "content": "Play!"}]}]]
)
producer._prompt_dl = [[{"prompt": [{"role": "user", "content": "Play!"}]}]]
# RepeatSampler pre-expands by num_generations=2.
_prompt_batch = [
{"prompt": [{"role": "user", "content": "Play!"}]},
{"prompt": [{"role": "user", "content": "Play!"}]},
]
producer._prompt_iter = iter([_prompt_batch])
producer._prompt_dl = [_prompt_batch]
result = producer.produce(model=MagicMock(), global_step=1)

View File

@@ -21,6 +21,51 @@ from unittest.mock import patch
import pytest
import torch
class TestPeftScatterMoELoRALayout:
"""CPU-only guards for PEFT target_parameters layout conversion."""
def test_peft_layout_keeps_a_and_reorders_b(self):
from axolotl.integrations.kernels.libs.scattermoe_lora.lora_layout import (
peft_lora_to_scattermoe,
)
E, r, K, N = 3, 2, 5, 7
scaling = 2.0
peft_A = torch.randn(E * r, K)
peft_B = torch.randn(N, E * r)
smoe_A, smoe_B = peft_lora_to_scattermoe(peft_A, peft_B, E, r)
assert smoe_A is peft_A
assert smoe_A.shape == (E * r, K)
assert smoe_B.shape == (N, E * r)
A_r = peft_A.reshape(E, r, K)
B_r = peft_B.reshape(N, r, E)
delta_peft = torch.einsum("o r e, e r i -> e o i", B_r, A_r) * scaling
for e in range(E):
A_e = smoe_A[e * r : (e + 1) * r, :]
B_e = smoe_B[:, e * r : (e + 1) * r]
torch.testing.assert_close(scaling * (B_e @ A_e), delta_peft[e])
def test_swapped_layout_fails_before_kernel_dispatch(self):
from axolotl.integrations.kernels.libs.scattermoe_lora.lora_layout import (
validate_scattermoe_lora_shapes,
)
E, r, K, N = 3, 2, 5, 7
expert_weights = torch.empty(E, K, N)
with pytest.raises(ValueError, match="Invalid ScatterMoE LoRA layout"):
validate_scattermoe_lora_shapes(
expert_weights=expert_weights,
lora_A=torch.empty(E * r, N),
lora_B=torch.empty(K, E * r),
)
# ============================================================================
# 1. KernelsArgs: disable_mlp_kernel validator
# ============================================================================

View File

@@ -38,6 +38,30 @@ def _reference_norm_noscale(x, eps):
return norm(x)
def _reference_partial_norm_rope(x, weight, cos, sin, eps):
"""Reference: Gemma4RMSNorm over the full head_dim, then stock
``apply_rotary_pos_emb`` over the first ``cos.shape[-1]`` columns, with
the trailing columns passed through unchanged. Mirrors how Llama-style
partial rotary is layered on top of the stock RMSNorm + RoPE primitives.
"""
from transformers.models.gemma4.modeling_gemma4 import (
Gemma4RMSNorm,
apply_rotary_pos_emb,
)
D = x.shape[-1]
n_rot = cos.shape[-1]
norm = Gemma4RMSNorm(D, eps=eps).to(x.device, x.dtype)
norm.weight.data.copy_(weight)
normed = norm(x)
if n_rot == D:
return apply_rotary_pos_emb(normed, cos, sin, unsqueeze_dim=2)
x_rot = normed[..., :n_rot]
x_pass = normed[..., n_rot:]
rotated = apply_rotary_pos_emb(x_rot, cos, sin, unsqueeze_dim=2)
return torch.cat([rotated, x_pass], dim=-1)
@pytest.fixture(
params=[
(2, 64, 32, 256), # sliding window layer shape
@@ -194,6 +218,172 @@ class TestFusedRMSNormRoPEBackward:
assert w.grad.abs().sum() > 0, "w.grad is all zeros"
class TestFusedRMSNormRoPEPartialRotary:
"""Partial-rotary: cos/sin last dim is smaller than head_dim.
Compares against the original primitives (`Gemma4RMSNorm` +
`apply_rotary_pos_emb`) applied to the rotated slice with the trailing
columns passed through. Without the kernel fix this used to crash with
`RuntimeError: shape '[..., D]' is invalid for input of size B*S*n_rot`.
"""
@pytest.mark.parametrize(
"B,S,H,D,n_rot",
[
(2, 16, 4, 64, 32), # half rotary (Llama-style 0.5)
(2, 16, 4, 64, 16), # quarter rotary
(2, 32, 8, 128, 64), # half rotary, larger heads
(1, 8, 2, 256, 64), # 26B sliding-shape, 0.25 partial
(1, 8, 2, 64, 64), # n_rot == D: must still match full-rotary path
],
ids=["half_64", "quarter_64", "half_128", "quarter_256", "full_64"],
)
def test_forward_matches_reference(self, B, S, H, D, n_rot):
from axolotl.kernels.gemma4_fused_rope import fused_rms_norm_rope
eps = 1e-6
x = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
weight = torch.randn(D, device="cuda", dtype=torch.bfloat16)
cos = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
sin = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
y_ref = _reference_partial_norm_rope(x.clone(), weight, cos, sin, eps)
y_fused = fused_rms_norm_rope(x.clone(), weight, cos, sin, eps=eps)
assert y_fused.shape == y_ref.shape == (B, S, H, D)
cos_sim = torch.nn.functional.cosine_similarity(
y_ref.flatten().float(), y_fused.flatten().float(), dim=0
)
assert cos_sim > 0.999, (
f"partial rotary forward cosine_sim={cos_sim:.6f} "
f"(B={B},S={S},H={H},D={D},n_rot={n_rot})"
)
# The pass-through tail must equal the reference RMSNorm output bit-
# for-bit (any deviation would mean the kernel is touching it with a
# spurious rotation, which is the original bug class).
torch.testing.assert_close(
y_fused[..., n_rot:], y_ref[..., n_rot:], rtol=1e-2, atol=1e-2
)
@pytest.mark.parametrize(
"B,S,H,D,n_rot",
[(2, 16, 4, 64, 32), (1, 8, 2, 256, 64)],
ids=["half_64", "quarter_256"],
)
def test_x_grad_matches_reference(self, B, S, H, D, n_rot):
from axolotl.kernels.gemma4_fused_rope import fused_rms_norm_rope
eps = 1e-6
cos = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
sin = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
weight_init = torch.randn(D, device="cuda", dtype=torch.bfloat16)
x_data = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
# Reference backward via the original primitives
x_ref = x_data.clone().requires_grad_(True)
w_ref = weight_init.clone()
y_ref = _reference_partial_norm_rope(x_ref, w_ref, cos, sin, eps)
y_ref.sum().backward()
# Fused backward
x_fused = x_data.clone().requires_grad_(True)
w_fused = weight_init.clone().requires_grad_(True)
y_fused = fused_rms_norm_rope(x_fused, w_fused, cos, sin, eps=eps)
y_fused.sum().backward()
cos_sim_x = torch.nn.functional.cosine_similarity(
x_fused.grad.flatten().float(), x_ref.grad.flatten().float(), dim=0
)
assert cos_sim_x > 0.999, f"partial rotary x grad cosine_sim={cos_sim_x:.6f}"
@pytest.mark.parametrize(
"B,S,H,D,n_rot",
[(2, 16, 4, 64, 32), (1, 8, 2, 256, 64)],
ids=["half_64", "quarter_256"],
)
def test_weight_grad_matches_reference(self, B, S, H, D, n_rot):
from transformers.models.gemma4.modeling_gemma4 import Gemma4RMSNorm
from axolotl.kernels.gemma4_fused_rope import fused_rms_norm_rope
eps = 1e-6
cos = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
sin = torch.randn(B, S, n_rot, device="cuda", dtype=torch.bfloat16)
weight_init = torch.randn(D, device="cuda", dtype=torch.bfloat16)
x_data = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
# Reference: Gemma4RMSNorm whose .weight collects grads, then partial
# rotary applied to the rotated slice.
norm_ref = Gemma4RMSNorm(D, eps=eps).cuda().to(torch.bfloat16)
norm_ref.weight = torch.nn.Parameter(weight_init.clone())
normed = norm_ref(x_data)
from transformers.models.gemma4.modeling_gemma4 import apply_rotary_pos_emb
rotated = apply_rotary_pos_emb(normed[..., :n_rot], cos, sin, unsqueeze_dim=2)
y_ref = torch.cat([rotated, normed[..., n_rot:]], dim=-1)
y_ref.sum().backward()
w_fused = weight_init.clone().requires_grad_(True)
fused_rms_norm_rope(x_data.clone(), w_fused, cos, sin, eps=eps).sum().backward()
cos_sim_w = torch.nn.functional.cosine_similarity(
w_fused.grad.flatten().float(),
norm_ref.weight.grad.flatten().float(),
dim=0,
)
assert cos_sim_w > 0.995, (
f"partial rotary weight grad cosine_sim={cos_sim_w:.6f}"
)
def test_full_rotary_unchanged_when_n_rot_equals_d(self):
"""Regression: passing cos/sin with shape == head_dim must still
match the full-rotary reference (the partial-rotary code path must
not perturb the existing full-rotary output)."""
from axolotl.kernels.gemma4_fused_rope import fused_rms_norm_rope
B, S, H, D = 2, 16, 4, 64
eps = 1e-6
x = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
weight = torch.randn(D, device="cuda", dtype=torch.bfloat16)
cos = torch.randn(B, S, D, device="cuda", dtype=torch.bfloat16)
sin = torch.randn(B, S, D, device="cuda", dtype=torch.bfloat16)
y_ref = _reference_norm_rope(x.clone(), weight, cos, sin, eps)
y_fused = fused_rms_norm_rope(x.clone(), weight, cos, sin, eps=eps)
cos_sim = torch.nn.functional.cosine_similarity(
y_ref.flatten().float(), y_fused.flatten().float(), dim=0
)
assert cos_sim > 0.999, f"full-rotary regression cos_sim={cos_sim:.6f}"
def test_validation_errors(self):
"""Wrapper rejects misshaped inputs cleanly (instead of a cryptic
Triton crash deeper in the kernel)."""
from axolotl.kernels.gemma4_fused_rope import fused_rms_norm_rope
B, S, H, D = 1, 4, 2, 64
x = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
w = torch.randn(D, device="cuda", dtype=torch.bfloat16)
# n_rot > head_dim
cos_big = torch.randn(B, S, D + 16, device="cuda", dtype=torch.bfloat16)
sin_big = torch.randn(B, S, D + 16, device="cuda", dtype=torch.bfloat16)
with pytest.raises(ValueError, match="cannot exceed head_dim"):
fused_rms_norm_rope(x, w, cos_big, sin_big)
# cos/sin last-dim mismatch
cos = torch.randn(B, S, 32, device="cuda", dtype=torch.bfloat16)
sin = torch.randn(B, S, 16, device="cuda", dtype=torch.bfloat16)
with pytest.raises(ValueError, match="same last dim"):
fused_rms_norm_rope(x, w, cos, sin)
# odd rotary dim
cos_odd = torch.randn(B, S, 31, device="cuda", dtype=torch.bfloat16)
sin_odd = torch.randn(B, S, 31, device="cuda", dtype=torch.bfloat16)
with pytest.raises(ValueError, match="must be even"):
fused_rms_norm_rope(x, w, cos_odd, sin_odd)
class TestFusedRMSNormNoScale:
"""Tests for v_norm (RMSNorm without learnable scale)."""

View File

@@ -0,0 +1,219 @@
"""Tests for the Gemma 4 fused-attention monkey-patch.
These tests exercise the patched ``Gemma4TextAttention.forward`` against
the stock implementation it replaces. The hybrid Gemma 4 model intentionally
mixes a sliding (`head_dim=32`) layer with a full-attention proportional-rope
layer (`global_head_dim=64`, `partial_rotary_factor=0.25`) so that the
partial-rotary RMSNorm+RoPE path through the fused Triton kernel is
exercised end-to-end (this is the bug originally documented in
``GEMMA4_FUSED_ROPE_HYBRID_ATTN_BUG.md``).
The full-model forward also pins that the fused forward keeps accepting
whatever call shape ``Gemma4TextDecoderLayer.forward`` produces in the
installed transformers version — so any future signature drift on
upstream's side trips a clear failure here instead of a confusing
TypeError deep in a training run.
"""
import pytest
import torch
pytestmark = [
pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required"),
]
pytest.importorskip(
"transformers.models.gemma4",
reason="fused_attn patch only matters when Gemma 4 is available",
)
@pytest.fixture
def restore_gemma4_attention():
"""Snapshot ``Gemma4TextAttention.forward`` and restore after the test
so the monkey-patch does not leak across the suite."""
from transformers.models.gemma4.modeling_gemma4 import Gemma4TextAttention
saved = Gemma4TextAttention.forward
yield Gemma4TextAttention
Gemma4TextAttention.forward = saved
def _build_hybrid_config():
"""Tiny hybrid Gemma 4 config: one sliding layer + one full-attention
layer with proportional rope and partial_rotary_factor=0.25. This is
the same shape pattern as ``google/gemma-4-26B-A4B-it`` but small
enough to fit on any GPU."""
from transformers.models.gemma4.configuration_gemma4 import Gemma4TextConfig
cfg = Gemma4TextConfig(
vocab_size=128,
hidden_size=64,
intermediate_size=128,
num_hidden_layers=2,
num_attention_heads=2,
num_key_value_heads=2,
head_dim=32,
global_head_dim=64,
layer_types=["sliding_attention", "full_attention"],
sliding_window=64,
max_position_embeddings=2048,
hidden_size_per_layer_input=16,
vocab_size_per_layer_input=128,
rope_parameters={
"sliding_attention": {
"rope_type": "default",
"rope_theta": 10000.0,
},
"full_attention": {
"rope_type": "proportional",
"rope_theta": 1000000.0,
"partial_rotary_factor": 0.25,
},
},
)
cfg._attn_implementation = "sdpa"
return cfg
def _build_model(seed=0):
from transformers.models.gemma4.modeling_gemma4 import Gemma4TextModel
torch.manual_seed(seed)
cfg = _build_hybrid_config()
return Gemma4TextModel(cfg).cuda().to(torch.bfloat16).eval()
class TestFusedAttnSignature:
"""The fused forward must accept the same call shape as
``Gemma4TextDecoderLayer`` produces in the installed transformers
version. Any signature drift surfaces here as a TypeError."""
def test_decoder_layer_can_call_fused_forward(self, restore_gemma4_attention):
"""Run a model forward that exercises the real
``Gemma4TextDecoderLayer -> Gemma4TextAttention`` call path with
the fused patch installed."""
from axolotl.monkeypatch.models.gemma4.fused_attn import (
patch_gemma4_fused_attn,
)
m = _build_model()
ids = torch.randint(0, 128, (2, 16), device="cuda")
mask = torch.ones(2, 16, dtype=torch.long, device="cuda")
patch_gemma4_fused_attn()
with torch.no_grad():
out = m(input_ids=ids, attention_mask=mask).last_hidden_state
assert out.shape == (2, 16, 64)
assert torch.isfinite(out).all()
class TestFusedAttnPerLayerCorrectness:
"""Compare the patched attention layer to the stock implementation
on a single forward call. This isolates the fused kernel correctness
from cross-layer numerical drift."""
def _run_attention(self, model, layer_idx, hidden_states, position_ids):
"""Call ``Gemma4TextAttention.forward`` (whatever is currently
installed) for one layer and return the output."""
attn = model.layers[layer_idx].self_attn
layer_type = model.config.layer_types[layer_idx]
cos, sin = model.rotary_emb(hidden_states, position_ids, layer_type)
out, _ = attn(
hidden_states=hidden_states,
position_embeddings=(cos, sin),
attention_mask=None,
shared_kv_states={},
)
return out
@pytest.mark.parametrize(
"layer_idx",
[0, 1],
ids=["sliding_head32", "global_head64_proportional"],
)
def test_forward_matches_stock(self, restore_gemma4_attention, layer_idx):
from axolotl.monkeypatch.models.gemma4.fused_attn import (
patch_gemma4_fused_attn,
)
m = _build_model(seed=1)
hs = torch.randn(2, 16, 64, device="cuda", dtype=torch.bfloat16)
pos = torch.arange(16, device="cuda").unsqueeze(0).expand(2, -1)
with torch.no_grad():
ref = self._run_attention(m, layer_idx, hs, pos)
patch_gemma4_fused_attn()
with torch.no_grad():
got = self._run_attention(m, layer_idx, hs, pos)
assert got.shape == ref.shape
assert torch.isfinite(got).all()
cos_sim = torch.nn.functional.cosine_similarity(
ref.flatten().float(), got.flatten().float(), dim=0
)
assert cos_sim > 0.999, (
f"layer {layer_idx} fused vs stock cosine_sim={cos_sim:.6f}"
)
# bf16 precision: a few millis of absolute drift per element is
# acceptable for a Q/K/V projection pipeline. Anything larger is
# a real bug.
torch.testing.assert_close(got, ref, rtol=5e-2, atol=5e-2)
class TestFusedAttnFullModel:
"""End-to-end model forward + backward through both layer types."""
def test_full_forward_matches_stock(self, restore_gemma4_attention):
from axolotl.monkeypatch.models.gemma4.fused_attn import (
patch_gemma4_fused_attn,
)
m = _build_model(seed=2)
ids = torch.randint(0, 128, (2, 32), device="cuda")
mask = torch.ones(2, 32, dtype=torch.long, device="cuda")
with torch.no_grad():
ref = m(input_ids=ids, attention_mask=mask).last_hidden_state.clone()
patch_gemma4_fused_attn()
with torch.no_grad():
got = m(input_ids=ids, attention_mask=mask).last_hidden_state.clone()
assert got.shape == ref.shape
assert torch.isfinite(got).all()
cos_sim = torch.nn.functional.cosine_similarity(
ref.flatten().float(), got.flatten().float(), dim=0
)
# End-to-end through 2 layers (RMSNorm, attention, MLP/MoE) in bf16
# accumulates a small amount of numerical drift; we just want to
# pin that the two paths are computing the same function.
assert cos_sim > 0.999, f"end-to-end cosine_sim={cos_sim:.6f}"
def test_backward_grad_flows_through_fused_path(self, restore_gemma4_attention):
"""Gradients must propagate through the fused RMSNorm+RoPE kernels
for both the sliding and proportional-rope layers."""
from axolotl.monkeypatch.models.gemma4.fused_attn import (
patch_gemma4_fused_attn,
)
m = _build_model(seed=3).train()
patch_gemma4_fused_attn()
ids = torch.randint(0, 128, (2, 16), device="cuda")
mask = torch.ones(2, 16, dtype=torch.long, device="cuda")
out = m(input_ids=ids, attention_mask=mask).last_hidden_state
out.sum().backward()
# Both layers must accumulate gradients on q_norm.weight and
# k_norm.weight — that proves the fused kernel ran the backward.
for i, layer in enumerate(m.layers[:2]):
attn = layer.self_attn
assert attn.q_norm.weight.grad is not None, f"layer {i} q_norm no grad"
assert attn.k_norm.weight.grad is not None, f"layer {i} k_norm no grad"
assert attn.q_norm.weight.grad.isfinite().all()
assert attn.k_norm.weight.grad.isfinite().all()
assert attn.q_norm.weight.grad.abs().sum() > 0
assert attn.k_norm.weight.grad.abs().sum() > 0

View File

@@ -0,0 +1,343 @@
"""Tests for the Gemma 4 hybrid-attention mask fix.
These tests pin the single critical behavior: after installing the patch,
``modeling_gemma4.create_causal_mask`` passes an SDPA-overridden config to
the underlying mask builder regardless of what the caller's config says.
This is what keeps full-attention (head_dim=512) global layers from
crashing at long sequence lengths — they need a 4D SDPA-format mask, not
the 2D FA2 mask that would be built from the model-level config.
The tests use a mocked ``create_causal_mask`` so they don't have to load
a real 26B Gemma 4 model or even have access to its weights. What matters
for the bug fix is which config is handed to the mask factory, not the
factory's actual output.
"""
from types import SimpleNamespace
from unittest.mock import MagicMock
import pytest
pytest.importorskip(
"transformers.models.gemma4",
reason="gemma4_hybrid_mask patch only matters when Gemma 4 is available",
)
@pytest.fixture
def restore_gemma4_module():
"""Snapshot ``modeling_gemma4.create_causal_mask`` and restore after
each test so patch state doesn't leak across the suite."""
from transformers.models.gemma4 import modeling_gemma4
saved = modeling_gemma4.create_causal_mask
yield modeling_gemma4
modeling_gemma4.create_causal_mask = saved
# Reset the module-level flag so the next test can re-install cleanly.
from axolotl.monkeypatch import gemma4_hybrid_mask
gemma4_hybrid_mask._PATCH_APPLIED = False
def test_patch_replaces_create_causal_mask(restore_gemma4_module):
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
original = modeling_gemma4.create_causal_mask
assert patch_gemma4_hybrid_mask() is True
assert modeling_gemma4.create_causal_mask is not original
assert modeling_gemma4.create_causal_mask._axolotl_original is original, (
"patched wrapper must expose the original reference for teardown"
)
def test_patch_is_idempotent(restore_gemma4_module):
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
patch_gemma4_hybrid_mask()
wrapper_first = modeling_gemma4.create_causal_mask
# Second call must not re-wrap the already-wrapped function (which
# would leak the original reference through a chain of wrappers).
patch_gemma4_hybrid_mask()
wrapper_second = modeling_gemma4.create_causal_mask
assert wrapper_first is wrapper_second
def test_patched_mask_forces_sdpa_config(restore_gemma4_module):
"""Core invariant: when the patched wrapper is called with a config
that says ``flash_attention_2``, the underlying mask factory receives
a shallow-copied config whose ``_attn_implementation`` is ``"sdpa"``.
Without this, the full-attention global layers get a 2D FA2 mask and
crash at long seq lens with the [B, H, S, S] / [B, S] expand error.
"""
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
# Swap in a mock BEFORE installing the patch so the wrapper captures
# it as the "original". The mock records every call so we can inspect
# what config got passed through.
mock_factory = MagicMock(name="create_causal_mask", return_value="mask_4d")
modeling_gemma4.create_causal_mask = mock_factory
patch_gemma4_hybrid_mask()
# Caller-supplied config says FA2 (that's the model-level setting).
caller_config = SimpleNamespace(
_attn_implementation="flash_attention_2",
head_dim=512,
some_other_attr="preserved",
)
result = modeling_gemma4.create_causal_mask(
caller_config,
inputs_embeds=None,
attention_mask=None,
past_key_values=None,
position_ids=None,
)
# Wrapper returned whatever the mock returned — no transformation of
# the result itself.
assert result == "mask_4d"
# The mock was called exactly once with a config whose
# ``_attn_implementation`` is sdpa, NOT the caller's fa2.
assert mock_factory.call_count == 1
(passed_config, *_), passed_kwargs = mock_factory.call_args
assert passed_config._attn_implementation == "sdpa"
# The wrapper must NOT mutate the caller's config in place — other
# mask builders (e.g. create_sliding_window_causal_mask) read from
# the same config and must still see fa2.
assert caller_config._attn_implementation == "flash_attention_2"
# Other attributes on the config must be preserved so the underlying
# factory has everything it needs (head_dim, rope_theta, vocab_size, ...).
assert passed_config.head_dim == 512
assert passed_config.some_other_attr == "preserved"
def test_patched_wrapper_passes_through_all_kwargs(restore_gemma4_module):
"""The wrapper must forward positional + keyword args to the original
unchanged, so transformers' own call-site in Gemma4TextModel.forward
keeps working across minor transformers-version signature drift."""
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
mock_factory = MagicMock(return_value="mask")
modeling_gemma4.create_causal_mask = mock_factory
patch_gemma4_hybrid_mask()
caller_config = SimpleNamespace(_attn_implementation="flash_attention_2")
modeling_gemma4.create_causal_mask(
caller_config,
"positional_arg",
inputs_embeds="embeds",
attention_mask="mask_2d",
past_key_values="cache",
position_ids="positions",
or_mask_function="or_fn",
)
args, kwargs = mock_factory.call_args
# First positional (after config override) is preserved.
assert args[1] == "positional_arg"
# All kwargs are forwarded untouched.
assert kwargs["inputs_embeds"] == "embeds"
assert kwargs["attention_mask"] == "mask_2d"
assert kwargs["past_key_values"] == "cache"
assert kwargs["position_ids"] == "positions"
assert kwargs["or_mask_function"] == "or_fn"
def test_unpatch_restores_original(restore_gemma4_module):
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import (
patch_gemma4_hybrid_mask,
unpatch_gemma4_hybrid_mask,
)
sentinel = MagicMock(name="original")
modeling_gemma4.create_causal_mask = sentinel
patch_gemma4_hybrid_mask()
assert modeling_gemma4.create_causal_mask is not sentinel
unpatch_gemma4_hybrid_mask()
assert modeling_gemma4.create_causal_mask is sentinel
def test_unpatch_is_safe_without_prior_patch(restore_gemma4_module):
from axolotl.monkeypatch.gemma4_hybrid_mask import unpatch_gemma4_hybrid_mask
# Should be a no-op, no exception.
unpatch_gemma4_hybrid_mask()
def test_sliding_window_mask_builder_is_not_patched(restore_gemma4_module):
"""Only ``create_causal_mask`` is overridden — the sliding-window
factory must remain bound to its original to preserve FA2 masks for
the sliding-attention layers. If we accidentally patch both, the
sliding layers get SDPA format and lose the FA2 speedup."""
modeling_gemma4 = restore_gemma4_module
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
if not hasattr(modeling_gemma4, "create_sliding_window_causal_mask"):
pytest.skip("transformers version has no create_sliding_window_causal_mask")
sliding_before = modeling_gemma4.create_sliding_window_causal_mask
patch_gemma4_hybrid_mask()
sliding_after = modeling_gemma4.create_sliding_window_causal_mask
assert sliding_after is sliding_before
# ---------------------------------------------------------------------------
# Integration tests with a tiny randomly-initialized Gemma4TextModel.
#
# These do NOT load real 26B weights. They build a ~350k-param Gemma 4 text
# model with 2 layers (one sliding, one full_attention), apply the hybrid
# attention path end-to-end, and run a forward pass with a padded
# attention_mask at a long-ish seq len. The invariant we're pinning is that
# the full_attention layer does not crash with the
# "Target sizes: [B, H, S, S]. Tensor sizes: [B, S]"
# error — the exact failure that blew up the Gemma 4 MoE 26B pilot at ~7k
# tokens in the FSDP2 training run.
# ---------------------------------------------------------------------------
def _build_tiny_gemma4_text_model():
"""Return a tiny randomly-initialized Gemma4TextModel with mixed layers."""
import torch
from transformers.models.gemma4.configuration_gemma4 import Gemma4TextConfig
from transformers.models.gemma4.modeling_gemma4 import Gemma4TextModel
cfg = Gemma4TextConfig(
vocab_size=128,
hidden_size=64,
intermediate_size=128,
num_hidden_layers=2,
num_attention_heads=2,
num_key_value_heads=2,
head_dim=32,
layer_types=["sliding_attention", "full_attention"],
sliding_window=64,
max_position_embeddings=2048,
hidden_size_per_layer_input=16,
vocab_size_per_layer_input=128,
)
# Caller-supplied attn impl simulates the pilot config (fa2 at model
# level). The hybrid patch is what makes this survive long context.
cfg._attn_implementation = "sdpa" # start safe; the test toggles fa2 later
torch.manual_seed(42)
model = Gemma4TextModel(cfg).eval()
return model, cfg
def _apply_hybrid_attn_inline(model, cfg):
"""Replicate what ``patch_manager._apply_gemma_hybrid_attention`` does
to a model, without needing a full PatchManager / pydantic cfg."""
import copy
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
for layer_idx, layer in enumerate(model.layers):
if cfg.layer_types[layer_idx] != "sliding_attention":
attn = getattr(layer, "self_attn", None)
if attn is not None and hasattr(attn, "config"):
sdpa_cfg = copy.copy(attn.config)
sdpa_cfg._attn_implementation = "sdpa"
attn.config = sdpa_cfg
patch_gemma4_hybrid_mask()
def test_tiny_gemma4_long_context_forward_does_not_crash(restore_gemma4_module):
"""End-to-end invariant: with the hybrid attn patch applied, a tiny
Gemma4TextModel runs a forward at long context (1024 tokens) with
real padding in the attention mask, producing the expected output
shape. This exercises the actual code path that crashed the pilot
without needing a real 26B checkpoint or CUDA."""
import torch
model, cfg = _build_tiny_gemma4_text_model()
_apply_hybrid_attn_inline(model, cfg)
B, S = 2, 1024
input_ids = torch.randint(0, cfg.vocab_size, (B, S))
attn_mask = torch.ones(B, S, dtype=torch.long)
# Pad positions in the second row. Without padding, SDPA falls back to
# ``is_causal=True`` with ``mask=None`` — we need a materialized 4D
# mask to exercise the actual bug site.
attn_mask[1, S // 2 :] = 0
with torch.no_grad():
out = model(input_ids=input_ids, attention_mask=attn_mask)
assert out.last_hidden_state.shape == (B, S, cfg.hidden_size)
assert torch.isfinite(out.last_hidden_state).all()
def test_patched_create_causal_mask_returns_4d_for_real_config(
restore_gemma4_module,
):
"""Hit the REAL ``create_causal_mask`` (not a mock) via the wrapper
and verify the returned mask is a 4D tensor — which is the shape the
SDPA-patched global layers need. Without the patch and with a
caller-supplied FA2 config this would return a 2D mask and the layer
would crash at long context."""
import torch
from transformers.cache_utils import DynamicCache
from transformers.models.gemma4.configuration_gemma4 import Gemma4TextConfig
from axolotl.monkeypatch.gemma4_hybrid_mask import patch_gemma4_hybrid_mask
patch_gemma4_hybrid_mask()
modeling_gemma4 = restore_gemma4_module
cfg = Gemma4TextConfig(
vocab_size=128,
hidden_size=64,
num_hidden_layers=2,
num_attention_heads=2,
num_key_value_heads=2,
head_dim=32,
layer_types=["sliding_attention", "full_attention"],
sliding_window=64,
max_position_embeddings=2048,
hidden_size_per_layer_input=16,
vocab_size_per_layer_input=128,
)
# Simulate the pilot: caller says flash_attention_2, but global layers
# were switched to SDPA per-layer. Without the patch, create_causal_mask
# would return an FA2 2D mask here and the SDPA layer would crash.
cfg._attn_implementation = "flash_attention_2"
B, S = 2, 1024
inputs_embeds = torch.zeros((B, S, cfg.hidden_size), dtype=torch.float32)
attention_mask = torch.ones((B, S), dtype=torch.long)
attention_mask[1, S // 2 :] = 0 # force the 4D materialized path
position_ids = torch.arange(S).unsqueeze(0).expand(B, -1)
past_key_values = DynamicCache(config=cfg)
mask = modeling_gemma4.create_causal_mask(
config=cfg,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
past_key_values=past_key_values,
position_ids=position_ids,
)
assert mask is not None
assert isinstance(mask, torch.Tensor)
assert mask.dim() == 4, (
f"expected a 4D SDPA-format mask, got {mask.dim()}D "
f"shape={tuple(mask.shape)}. The full_attention global layers need "
"this shape or they crash at long context."
)
assert mask.shape[0] == B
assert mask.shape[-1] == S
assert mask.shape[-2] == S
# Caller's config must be untouched — other code paths still read it.
assert cfg._attn_implementation == "flash_attention_2"

View File

@@ -491,7 +491,8 @@ class TestEfficientMerge:
out_features = 4
alpha = 4
base = torch.randn(num_experts, in_features, out_features)
# PEFT ParamWrapper treats non-transposed 3D weights as (experts, out, in)
base = torch.randn(num_experts, out_features, in_features)
lora_a = torch.randn(r * num_experts, in_features)
lora_b = torch.randn(out_features, r * num_experts)
@@ -507,7 +508,7 @@ class TestEfficientMerge:
scale = alpha / r
wa = lora_a.reshape(num_experts, r, in_features)
wb = lora_b.reshape(out_features, r, num_experts)
manual_delta = torch.einsum("o r e, e r i -> e i o", wb, wa) * scale
manual_delta = torch.einsum("o r e, e r i -> e o i", wb, wa) * scale
for e in range(num_experts):
assert torch.allclose(merged[e], base[e] + manual_delta[e], atol=1e-5), (
f"Expert {e} mismatch"

View File

@@ -5,6 +5,8 @@ Covers:
- save_strategy: 'best' requires metric_for_best_model
- streaming=True with val_set_size > 0 is rejected
- lora_target_modules with invalid regex patterns is rejected
- GRPO: generation batch size must be divisible by num_generations,
num_generations >= 2, and effective_gbs >= num_generations * world_size
"""
import pytest
@@ -117,3 +119,136 @@ class TestLoraTargetModulesRegexValidator:
)
with pytest.raises(ValueError, match="invalid regex pattern"):
validate_config(cfg)
class TestGRPOBatchSizeValidator:
"""GRPO requires (mb*GA) % num_generations == 0 and num_generations >= 2.
These call the @model_validator(mode="before") classmethod directly on a
plain dict — same input shape it receives during full Pydantic validation,
just without dragging in unrelated fields (datasets / model loading / etc.)
that aren't relevant to what's under test. The validator is registered on
``RLValidationMixin`` (which ``AxolotlInputConfig`` inherits) so this is the
same code path ``axolotl train`` exercises.
"""
@staticmethod
def _check(data):
from axolotl.utils.schemas.validation import RLValidationMixin
return RLValidationMixin.check_grpo_batch_size_divisibility(data)
def test_divisible_passes(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"trl": {"num_generations": 4},
}
# Should return data unchanged (no exception)
out = self._check(data)
assert out["trl"]["num_generations"] == 4
def test_non_divisible_raises(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 2,
"trl": {"num_generations": 4},
}
with pytest.raises(ValueError, match="num_generations"):
self._check(data)
def test_non_divisible_error_includes_fix_hint(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 3,
"trl": {"num_generations": 4},
}
with pytest.raises(ValueError, match="gradient_accumulation_steps: 4"):
self._check(data)
def test_num_generations_one_raises(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"trl": {"num_generations": 1},
}
with pytest.raises(ValueError, match=r"num_generations >= 2"):
self._check(data)
def test_explicit_generation_batch_size_divisible_passes(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 1,
"trl": {"num_generations": 4, "generation_batch_size": 8},
}
out = self._check(data)
assert out["trl"]["generation_batch_size"] == 8
def test_explicit_generation_batch_size_non_divisible_raises(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 1,
"trl": {"num_generations": 4, "generation_batch_size": 6},
}
with pytest.raises(ValueError, match="trl.generation_batch_size"):
self._check(data)
def test_non_grpo_skips_check(self):
# Anything other than rl=grpo should pass through untouched, even
# with non-divisible batch sizes — they're irrelevant to other RL
# methods that don't use group-relative advantages.
data = {
"rl": "dpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 3,
"trl": {"num_generations": 4},
}
assert self._check(data) is data
def test_no_rl_set_skips_check(self):
data = {
"micro_batch_size": 1,
"gradient_accumulation_steps": 3,
}
assert self._check(data) is data
def test_grpo_without_num_generations_skips_check(self):
# If num_generations isn't set, TRL uses its own default — we don't
# have enough info to validate, so the validator must short-circuit
# rather than guess.
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 3,
"trl": {},
}
out = self._check(data)
assert out["rl"] == "grpo"
def test_multi_rank_group_size_check(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 4, # gbs=4
"world_size": 2, # need gbs >= 4*2 = 8
"trl": {"num_generations": 4},
}
with pytest.raises(ValueError, match=r"world_size=2"):
self._check(data)
def test_multi_rank_group_size_satisfied(self):
data = {
"rl": "grpo",
"micro_batch_size": 1,
"gradient_accumulation_steps": 8, # gbs=8 >= 4*2
"world_size": 2,
"trl": {"num_generations": 4},
}
out = self._check(data)
assert out["gradient_accumulation_steps"] == 8