Skip to content

Add ROCm support for CI test_qwen3_4B_ckpt #1621

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 1 commit into
base: main
Choose a base branch
from
Open

Add ROCm support for CI test_qwen3_4B_ckpt #1621

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
4 changes: 3 additions & 1 deletion miles/backends/megatron_utils/ci_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,11 +169,13 @@ def check_peak_gpu_memory_after_load(args) -> None:
return

# Threshold 20 GB is midpoint between ~16.9 GB (with) and ~22.4 GB (without) on 8xH200.
# ROCm async 4+4 config uses TP=1,CP=1 (full model per GPU), so ~34 GB is expected.
peak_gpu_gb = torch.cuda.max_memory_allocated() / (1024**3)
rank = torch.distributed.get_rank() if torch.distributed.is_initialized() else 0
logger.info(f"[CI low-memory-resume] Rank {rank} peak GPU memory: {peak_gpu_gb:.2f} GB")

threshold_gb = 20.0
is_rocm = hasattr(torch.version, "hip") and torch.version.hip is not None
threshold_gb = 40.0 if is_rocm else 20.0
assert peak_gpu_gb < threshold_gb, (
f"[Rank {rank}] Peak GPU memory ({peak_gpu_gb:.2f} GB) exceeds threshold ({threshold_gb} GB). "
f"--low-memory-resume optimization may not be working correctly."
Expand Down
4 changes: 3 additions & 1 deletion miles/backends/training_utils/ci_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,9 @@ def check_kl(args: Namespace, log_dict: dict[str, float], step_id: int, accumula
# small floating-point differences, so use a relaxed threshold.
assert abs(log_dict["train/ppo_kl"]) < 1e-8 and abs(log_dict["train/pg_clipfrac"]) < 1e-10, f"{log_dict=}"
else:
assert abs(log_dict["train/ppo_kl"]) < 1e-9 and abs(log_dict["train/pg_clipfrac"]) < 1e-10, f"{log_dict=}"
_is_rocm = hasattr(torch.version, "hip") and torch.version.hip is not None
_tol = 1e-8 if _is_rocm else 1e-10
assert abs(log_dict["train/ppo_kl"]) < _tol and abs(log_dict["train/pg_clipfrac"]) < _tol, f"{log_dict=}"
if accumulated_step_id == 0 and "train/kl_loss" in log_dict and not args.use_rollout_routing_replay:
assert abs(log_dict["train/kl_loss"]) < 1e-9, f"{log_dict=}"

Expand Down
19 changes: 19 additions & 0 deletions miles/utils/external_utils/command_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,11 +45,18 @@ def convert_checkpoint(
"--master-addr {{master_addr}} " "--master-port 23456 " "--nnodes={{nnodes}} " "--node-rank {{node_rank}} "
)

rocm_env = ""
if os.path.exists("/opt/rocm"):
for var in ("NVTE_FUSED_ATTN", "NVTE_FLASH_ATTN", "NVTE_UNFUSED_ATTN"):
os.environ.pop(var, None)
rocm_env = "unset NVTE_FUSED_ATTN NVTE_FLASH_ATTN NVTE_UNFUSED_ATTN && "

if multinode:
fn = partial(exec_command_all_ray_node, num_nodes=num_nodes)
else:
fn = exec_command
fn(
f"{rocm_env}"
f"source {repo_base_dir}/scripts/models/{megatron_model_type}.sh && "
f"PYTHONPATH={megatron_path} "
f"torchrun "
Expand Down Expand Up @@ -133,6 +140,18 @@ def execute_train(
"true; "
)

rocm_env = ""
if os.path.exists("/opt/rocm"):
for var in ("NVTE_FUSED_ATTN", "NVTE_FLASH_ATTN", "NVTE_UNFUSED_ATTN"):
os.environ.pop(var, None)
os.environ.setdefault("RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES", "1")
os.environ.setdefault("HIP_VISIBLE_DEVICES", ",".join(str(i) for i in range(num_gpus_per_node)))
rocm_env = (
f"unset NVTE_FUSED_ATTN NVTE_FLASH_ATTN NVTE_UNFUSED_ATTN && "
f"export RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES=1 && "
f"export HIP_VISIBLE_DEVICES={os.environ['HIP_VISIBLE_DEVICES']} && "
)

if not external_ray:
exec_command(
# will prevent ray from buffering stdout/stderr
Expand Down
126 changes: 91 additions & 35 deletions tests/e2e/ckpt/test_qwen3_4B_ckpt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,8 @@
register_rocm_ci(est_time=1200, suite="stage-c-8-gpu-mi350", labels=["ckpt"])

ENABLE_EVAL = bool(int(os.environ.get("MILES_TEST_ENABLE_EVAL", "1")))
TIGHT_HOST_MEMORY = bool(int(os.environ.get("MILES_TEST_TIGHT_HOST_MEMORY", "1")))
IS_ROCM = os.path.exists("/opt/rocm")

MODEL_NAME = "Qwen3-4B"
MODEL_TYPE = "qwen3-4B"
Expand Down Expand Up @@ -66,17 +68,33 @@ def execute(mode: str = "", ckpt_step: int | None = None):
"--balance-data "
)

perf_args = (
"--tensor-model-parallel-size 2 "
"--sequence-parallel "
"--pipeline-model-parallel-size 2 "
"--context-parallel-size 2 "
"--recompute-granularity full "
"--recompute-method uniform "
"--recompute-num-layers 1 "
"--use-dynamic-batch-size "
"--max-tokens-per-gpu 16384 "
)
if IS_ROCM:
# ROCm async 4+4: TP=1, CP=1 (full model per GPU)
actor_gpus = 4
perf_args = (
"--tensor-model-parallel-size 1 "
"--sequence-parallel "
"--pipeline-model-parallel-size 1 "
"--context-parallel-size 1 "
"--recompute-granularity full "
"--recompute-method uniform "
"--recompute-num-layers 1 "
"--use-dynamic-batch-size "
"--max-tokens-per-gpu 2048 "
)
else:
actor_gpus = NUM_GPUS
perf_args = (
"--tensor-model-parallel-size 2 "
"--sequence-parallel "
"--pipeline-model-parallel-size 1 "
"--context-parallel-size 2 "
"--recompute-granularity full "
"--recompute-method uniform "
"--recompute-num-layers 1 "
"--use-dynamic-batch-size "
f"--max-tokens-per-gpu {2048 if TIGHT_HOST_MEMORY else 16384} "
)

ppo_args = (
"--advantage-estimator grpo "
Expand All @@ -87,36 +105,73 @@ def execute(mode: str = "", ckpt_step: int | None = None):
"--eps-clip 0.2 "
)

optimizer_args = (
"--optimizer adam "
"--lr 1e-6 "
"--lr-decay-style constant "
"--weight-decay 0.1 "
"--adam-beta1 0.9 "
"--adam-beta2 0.98 "
)

sglang_args = "--rollout-num-gpus-per-engine 2 --sglang-mem-fraction-static 0.7 --sglang-cuda-graph-bs 1 2 4 8 16 "
if IS_ROCM:
optimizer_args = (
"--optimizer adam "
"--lr 1e-6 "
"--lr-decay-style constant "
"--weight-decay 0.1 "
"--adam-beta1 0.9 "
"--adam-beta2 0.98 "
"--optimizer-cpu-offload "
"--overlap-cpu-optimizer-d2h-h2d "
"--use-precision-aware-optimizer "
)
else:
optimizer_args = (
"--optimizer adam "
"--lr 1e-6 "
"--lr-decay-style constant "
"--weight-decay 0.1 "
"--adam-beta1 0.9 "
"--adam-beta2 0.98 "
)

if IS_ROCM:
rollout_gpus = 4
sglang_args = (
f"--rollout-num-gpus {rollout_gpus} "
"--rollout-num-gpus-per-engine 2 "
"--sglang-mem-fraction-static 0.8 "
"--sglang-disable-custom-all-reduce "
)
else:
sglang_args = "--rollout-num-gpus-per-engine 2 --sglang-mem-fraction-static 0.8 --sglang-cuda-graph-bs 1 2 4 8 16 "

ci_args = "--ci-test "
if mode in {"save", "async_save"}:
ci_args += "--ci-save-model-hash "
if mode == "load":
ci_args += "--ci-check-model-hash "

misc_args = (
# default dropout in megatron is 0.1
"--attention-dropout 0.0 "
"--hidden-dropout 0.0 "
# should be good for model performance
"--accumulate-allreduce-grads-in-fp32 "
"--attention-softmax-in-fp32 "
# need to comment this when using model with MLA
"--attention-backend flash "
"--actor-num-nodes 1 "
f"--actor-num-gpus-per-node {NUM_GPUS} "
"--colocate "
)
if IS_ROCM:
misc_args = (
"--attention-dropout 0.0 "
"--hidden-dropout 0.0 "
"--accumulate-allreduce-grads-in-fp32 "
"--attention-softmax-in-fp32 "
"--attention-backend auto "
"--update-weights-interval 2 "
"--actor-num-nodes 1 "
f"--actor-num-gpus-per-node {actor_gpus} "
)
else:
misc_args = (
"--attention-dropout 0.0 "
"--hidden-dropout 0.0 "
"--accumulate-allreduce-grads-in-fp32 "
"--attention-softmax-in-fp32 "
"--attention-backend flash "
"--actor-num-nodes 1 "
"--actor-num-gpus-per-node 8 "
"--colocate "
)

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"

train_args = (
f"{ckpt_args} "
Expand All @@ -134,7 +189,8 @@ def execute(mode: str = "", ckpt_step: int | None = None):
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