Skip to content

tests/mimo-7B: switch ROCm path to async 4+4 to avoid 8-rank RCCL hang and no gradient accumulation fusion #1615

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sreerohi wants to merge 2 commits into
base: main
Choose a base branch
from
+37 −10
Open

tests/mimo-7B: switch ROCm path to async 4+4 to avoid 8-rank RCCL hang and no gradient accumulation fusion #1615

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a52a0bb
tests/mimo-7B: switch ROCm path to async 4+4 to avoid 8-rank RCCL hang
sreerohi May 21, 2026
07af920
tests/mimo-7B: import torch and guard hip attr for ROCm GPU-count check
Jul 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 37 additions & 10 deletions tests/e2e/megatron/test_mimo_7B_mtp_only_grad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,8 @@

import os

import torch

from tests.ci.ci_register import register_cuda_ci, register_rocm_ci

import miles.utils.external_utils.command_utils as U
Expand All @@ -20,7 +22,9 @@
MODEL_NAME = "MiMo-7B-RL"
MODEL_TYPE = "mimo-7B-rl"
NUM_GPUS = 4

IS_ROCM = hasattr(torch.version, "hip") and torch.version.hip is not None
if IS_ROCM:
NUM_GPUS = 8

def prepare():
"""Download model and convert checkpoint with MTP layers."""
Expand Down Expand Up @@ -60,8 +64,11 @@ def execute():
"--global-batch-size 8 "
)

actor_gpus = 4 if IS_ROCM else NUM_GPUS
rollout_gpus = 4 if IS_ROCM else NUM_GPUS

perf_args = (
"--tensor-model-parallel-size 2 "
f"--tensor-model-parallel-size {1 if IS_ROCM else 2} "
"--sequence-parallel "
"--pipeline-model-parallel-size 1 "
"--context-parallel-size 1 "
Expand Down Expand Up @@ -89,17 +96,27 @@ def execute():
"--weight-decay 0.1 "
"--adam-beta1 0.9 "
"--adam-beta2 0.98 "
) + (
"--optimizer-cpu-offload "
"--overlap-cpu-optimizer-d2h-h2d "
"--use-precision-aware-optimizer " if IS_ROCM else ""
)

sglang_args = (
"--rollout-num-gpus-per-engine 1 "
"--sglang-mem-fraction-static 0.7 "
"--sglang-enable-metrics "
(
f"--rollout-num-gpus {rollout_gpus} "
"--rollout-num-gpus-per-engine 2 "
"--sglang-mem-fraction-static 0.8 "
if IS_ROCM
else "--rollout-num-gpus-per-engine 1 "
"--sglang-mem-fraction-static 0.7 "
)
+ "--sglang-enable-metrics "
"--sglang-speculative-algorithm EAGLE "
"--sglang-speculative-num-steps 2 "
"--sglang-speculative-eagle-topk 1 "
"--sglang-speculative-num-draft-tokens 3 "
)
) + ("--sglang-disable-custom-all-reduce " if IS_ROCM else "")

# Enable MTP training with loss scaling
mtp_args = "--mtp-num-layers 1 " "--enable-mtp-training " "--mtp-loss-scaling-factor 0.2 "
Expand All @@ -115,10 +132,13 @@ def execute():
"--hidden-dropout 0.0 "
"--accumulate-allreduce-grads-in-fp32 "
"--attention-softmax-in-fp32 "
"--attention-backend flash "
f"--attention-backend {'auto' if IS_ROCM else 'flash'} "
"--actor-num-nodes 1 "
f"--actor-num-gpus-per-node {NUM_GPUS} "
"--colocate "
f"--actor-num-gpus-per-node {actor_gpus if IS_ROCM else NUM_GPUS} "
) + (
"--update-weights-interval 2 --no-gradient-accumulation-fusion "
if IS_ROCM
else "--colocate "
)

train_args = (
Expand All @@ -134,11 +154,18 @@ def execute():
f"{misc_args} "
)

extra_env_vars = {"MILES_EXPERIMENTAL_ROLLOUT_REFACTOR": "1"}
if IS_ROCM:
extra_env_vars["SGLANG_SET_CPU_AFFINITY"] = "0"
extra_env_vars["RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES"] = "1"
extra_env_vars["HIP_VISIBLE_DEVICES"] = "0,1,2,3,4,5,6,7"

U.execute_train(
train_args=train_args,
num_gpus_per_node=NUM_GPUS,
megatron_model_type=MODEL_TYPE,
extra_env_vars={"MILES_EXPERIMENTAL_ROLLOUT_REFACTOR": "1"},
train_script="train_async.py" if IS_ROCM else "train.py",
extra_env_vars=extra_env_vars,
)


Expand Down