Skip to content

[skyrl-train] Update Megatron + LoRA to use PEFT only bridge #885

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

Draft
erictang000 wants to merge 9 commits into
base: main
Choose a base branch
from
Draft

[skyrl-train] Update Megatron + LoRA to use PEFT only bridge #885

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
15f6993
x
erictang000 Jan 15, 2026
56ae05b
x
erictang000 Jan 15, 2026
52015fa
x
erictang000 Jan 15, 2026
5d39788
x
erictang000 Jan 15, 2026
262e6c9
fix non-megatron gpu tests
erictang000 Jan 16, 2026
3708539
bump te to 2.10 to fix triton 3.5 + te2.9 incompatibility
erictang000 Jan 16, 2026
aa5b00e
Merge branch 'main' of https://github.com/erictang000/SkyRL into upgr…
erictang000 Jan 16, 2026
991ebce
Merge branch 'upgrade_vllm' of https://github.com/erictang000/SkyRL i…
erictang000 Jan 16, 2026
88fd6c3
x
erictang000 Jan 17, 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
25 changes: 16 additions & 9 deletions skyrl-train/pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ override-dependencies = [
"nvidia-resiliency-ext; sys_platform == 'never'",
"mamba-ssm; sys_platform == 'never'",
"causal-conv1d; sys_platform == 'never'",
"transformer-engine[pytorch]==2.9.0",
"transformer-engine[pytorch]==2.10.0",
"megatron-core==0.15.0"
]
[tool.uv.extra-build-dependencies]
Expand All @@ -100,7 +100,8 @@ flashinfer-jit-cache = { index = "flashinfer-cu128", marker = "extra == 'vllm' o
flashinfer-python = [
{ url = "https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl", marker = "extra == 'sglang' and extra != 'mcore' and extra != 'vllm'" }
]
megatron-bridge = {git = "https://github.com/NVIDIA-NeMo/Megatron-Bridge", rev = "953aabf75c0500180dc14a6a76cf9e7e7c4baec7"}
megatron-bridge = {git = "https://github.com/NVIDIA-NeMo/Megatron-Bridge", rev = "04e370eedf8cc44a812189a19f2171d90555c07a"}
vllm = { index = "vllm-nightly-4829148" }


[project.optional-dependencies]
Expand All @@ -125,9 +126,9 @@ sandboxes = [
"litellm[proxy]>=1.67.5",
]
vllm = [
"vllm==0.11.0",
"vllm", # Version resolved from vllm-nightly index (commit 4829148)
"flash-attn==2.8.3",
"torch==2.8.0",
"torch==2.9.1", # Required by vllm nightly
"flashinfer-python",
"flashinfer-jit-cache",
"torchvision"
Expand All @@ -140,15 +141,15 @@ sglang = [
"torchvision",
]
mcore = [
"transformer-engine[pytorch]==2.9.0",
"transformer-engine[pytorch]==2.10.0",
"flash-attn==2.8.1",
"vllm==0.11.0",
"torch==2.8.0",
"flashinfer-python==0.5.2",
"vllm", # Version resolved from vllm-nightly index (commit 4829148)
"torch==2.9.1", # Required by vllm nightly
"flashinfer-python==0.5.3",
"torchvision",
"megatron-bridge @ git+https://github.com/NVIDIA-NeMo/Megatron-Bridge.git@v0.2.0",
"megatron-core==0.15.0",
"flashinfer-jit-cache==0.5.2",
"flashinfer-jit-cache==0.5.3",
"nvidia-modelopt",
]
flashrl = [
Expand Down Expand Up @@ -176,6 +177,12 @@ name = "flashinfer-cu128"
url = "https://flashinfer.ai/whl/cu128"
explicit = true

[[tool.uv.index]]
name = "vllm-nightly-4829148"
# Replace cu130 with whatever variant you need (cu129, cu130, etc).
url = "https://wheels.vllm.ai/482914849cf9ce61d3e0dffaa35096bb34de58f5"
explicit = true

[tool.setuptools]
include-package-data = true

Expand Down
15 changes: 12 additions & 3 deletions skyrl-train/skyrl_train/inference_engines/vllm/vllm_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -364,27 +364,36 @@ def _create_engine(self, *args, **kwargs):
model_name = model_path

base_model_paths = [BaseModelPath(name=model_name, model_path=model_path)]
models = OpenAIServingModels(engine, model_config, base_model_paths)

# vllm >= 0.11.2 removed model_config from OpenAI serving APIs
is_new_api = version.parse(vllm.__version__) >= version.parse("0.11.2")
legacy_kwargs = {}
if is_new_api:
models = OpenAIServingModels(engine, base_model_paths)
else:
models = OpenAIServingModels(engine, model_config, base_model_paths)
legacy_kwargs["model_config"] = model_config

# TODO(Charlie): revisit kwargs `enable_auto_tools` and `tool_parser` when we need to
# support OAI-style tool calling; and `request_logger` for better debugging.
self.openai_serving_chat = OpenAIServingChat(
engine_client=engine,
model_config=model_config,
models=models,
response_role="assistant",
request_logger=None,
chat_template=None,
chat_template_content_format="auto",
**legacy_kwargs,
**openai_kwargs,
)

# TODO(Charlie): revisit kwargs `return_tokens_as_token_ids`,
# `enable_prompt_tokens_details`, `enable_force_include_usage`.
self.openai_serving_completion = OpenAIServingCompletion(
engine_client=engine,
model_config=model_config,
models=models,
request_logger=None,
**legacy_kwargs,
)
return engine

Expand Down
6 changes: 3 additions & 3 deletions skyrl-train/skyrl_train/inference_engines/vllm/vllm_server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,8 @@
import signal
import uvloop
from vllm import AsyncLLMEngine
from vllm.utils import FlexibleArgumentParser, set_ulimit
from vllm.utils.argparse_utils import FlexibleArgumentParser
from vllm.utils.system_utils import set_ulimit
from vllm.entrypoints.openai.cli_args import (
make_arg_parser,
validate_parsed_serve_args,
Expand Down Expand Up @@ -121,8 +122,7 @@ async def _destroy_weights_update_group(request: Request):
)
return {"status": "ok"}

vllm_config = await engine.get_vllm_config()
await init_app_state(engine, vllm_config, app.state, args)
await init_app_state(engine, app.state, args)

shutdown_task = await serve_http(
app,
Expand Down
9 changes: 9 additions & 0 deletions skyrl-train/skyrl_train/workers/megatron/megatron_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,6 +137,7 @@ def extract_weights(self, dtype: torch.dtype):
self.actor_module,
show_progress=False,
conversion_tasks=None,
merge_weights=False,
)

for name, tensor in hf_params_generator:
Expand All @@ -156,7 +157,9 @@ def extract_weights(self, dtype: torch.dtype):
self.actor_module,
show_progress=False,
conversion_tasks=bucket,
merge_weights=False,
)
breakpoint(0)

# Collect all parameters in this bucket into one chunk
names = []
Expand Down Expand Up @@ -647,6 +650,12 @@ async def broadcast_to_inference_engines(self, inference_engine_client):
torch.cuda.empty_cache()

# Extract and send weights using the sender created at init time
if self._is_lora:
# extract weights


#
return
await self._weight_transfer_sender.send_chunks(self.weight_extractor.extract_weights(generator_dtype))

if cache_reset_task is not None:
Expand Down
14 changes: 7 additions & 7 deletions skyrl-train/tests/gpu/gpu_ci/test_megatron_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,19 +111,19 @@ def get_test_training_batch(batch_size=4) -> TrainingInputBatch:


@pytest.mark.parametrize(
("colocate_all", "inference_tp", "megatron_tp", "megatron_pp", "megatron_ep", "megatron_etp", "lora"),
[(True, 4, 2, 2, 1, None, False), (False, 2, 2, 1, 1, None, False), (True, 4, 2, 2, 1, None, True)],
ids=["colocate_all", "non_colocated", "colocate_all_lora"],
("colocate_all", "inference_tp", "megatron_tp", "megatron_pp", "megatron_ep", "megatron_etp", "lora", "model_name"),
[(True, 4, 2, 2, 1, None, False, MODEL_NAME), (False, 2, 2, 1, 1, None, False, MODEL_NAME), (True, 4, 2, 2, 1, None, True, MODEL_NAME), (True, 4, 4, 1, 4, 1, True, MOE_MODEL_NAME)],
ids=["colocate_all_qwen3_0.6b", "non_colocated_qwen3_0.6b", "colocate_all_lora_qwen3_0.6b", "colocate_all_moe_lora_qwen3_30b_a3b"],
)
@pytest.mark.megatron
def test_megatron_policy_weight_sync(
colocate_all, inference_tp, megatron_tp, megatron_pp, megatron_ep, megatron_etp, lora
colocate_all, inference_tp, megatron_tp, megatron_pp, megatron_ep, megatron_etp, lora, model_name
):
"""
Test that we can sync weights between policy and inference for megatron then run inference
"""
try:
cfg = get_test_actor_config(model_name=MODEL_NAME)
cfg = get_test_actor_config(model_name=model_name)
if lora:
cfg.trainer.policy.model.lora.rank = 16
cfg.trainer.policy.model.lora.alpha = 16
Expand All @@ -141,7 +141,7 @@ def test_megatron_policy_weight_sync(

# If colocate is True, this will load the engine, sleep, and wake up the engine
client, pg = init_inference_engines(
model=MODEL_NAME,
model=model_name,
cfg=cfg,
use_local=True,
async_engine=cfg.generator.async_engine,
Expand Down Expand Up @@ -172,7 +172,7 @@ def test_megatron_policy_weight_sync(
policy.offload_to_cpu()
asyncio.run(client.wake_up(tags=["kv_cache"]))
sampling_params = get_sampling_params_for_backend(cfg.generator.backend, cfg.generator.sampling_params)
outputs = asyncio.run(run_inference(client, get_test_prompts(MODEL_NAME), sampling_params))
outputs = asyncio.run(run_inference(client, get_test_prompts(model_name), sampling_params))

print(f"Example output: {outputs['responses'][0]}, {outputs['stop_reasons'][0]}")
finally:
Expand Down
3 changes: 3 additions & 0 deletions skyrl-train/tests/gpu/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -445,6 +445,9 @@ def get_free_port():
# For standalone server, we use mp for now.
"--distributed-executor-backend",
"mp",
# vLLM 0.13+ V1 engine spawns worker processes that can't inherit CUDA context
# when CUDA_VISIBLE_DEVICES is set. Disable frontend multiprocessing to fix this.
"--disable-frontend-multiprocessing",
"--dtype",
"bfloat16",
"--host",
Expand Down
Loading
Loading