GPU Model Runner V2 #25266
Draft
GPU Model Runner V2 #25266
+2,360
−12
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
yinghai
reviewed
Sep 22, 2025
| # Request not found. | ||
| return | ||
| self.index_to_req_id.pop(req_idx, None) | ||
| self.free_indices.append(req_idx) |
There was a problem hiding this comment.
Would be easier to make this a set so that we can check uniqueness?
vllm/v1/worker/gpu/attn_utils.py
Outdated
| for group in kv_cache_config.kv_cache_groups: | ||
| for layer_name in group.layer_names: | ||
| layer_names.add(layer_name) | ||
| assert layer_names == set(kv_cache_raw_tensors.keys() |
jikunshang
reviewed
Sep 24, 2025
| attn_layers = get_layers_from_vllm_config(vllm_config, Attention) | ||
| for kv_cache_group_spec in kv_cache_config.kv_cache_groups: | ||
| layer_names = kv_cache_group_spec.layer_names | ||
| any_layer_name = next(iter(layer_names)) |
There was a problem hiding this comment.
This appears to assume an always on hybrid-kv-cache manager; I am 100% supportive of this but the iirc the reason we still supported disabling the hybrid-kv-cache manager was because P/D did not support the hybrid kv-cache manager yet; cc @NickLucche @heheda12345 do you know the state of P/D + hybrid-kv-cache? is there any other reason we would want to disable the hybrid-kv-cache?
There was a problem hiding this comment.
Still not supported on PD, @KuntaiDu has a series of PRs to enable hybrid allocator + kv connectors first
vllm/v1/worker/gpu/model_runner.py
Outdated
| [scheduler_output.num_scheduled_tokens[i] for i in req_ids], | ||
| dtype=np.int32) | ||
|
|
||
| # TODO(woosuk): Support CUDA graphs. |
There was a problem hiding this comment.
nice! happy to see this above attention metadata prep; we should definitely treat full-cudagraphs as first-class and I think padding for cudagraphs before attention metadata prep has been a major source of headaches/bugs; very excited about this!
jikunshang
reviewed
Sep 25, 2025
njhill
reviewed
Sep 29, 2025
There was a problem hiding this comment.
@WoosukKwon great work, I really like the direction.
Left a few comments of random things I noticed.
And of course there will be some work to integrate with other features as discussed.
vllm/v1/worker/gpu/states.py
Outdated
| def add_request( | ||
| self, | ||
| req_id: str, | ||
| prompt_token_ids: list[int], |
There was a problem hiding this comment.
Maybe we can also look at keeping things as np arrays end to end, or array.array for lists that grow (for later).
| ) | ||
| self.np = np.zeros_like(self.buffer.np) | ||
|
|
||
| def copy_np_to_gpu(self, x: np.ndarray) -> torch.Tensor: |
| # TODO(woosuk): Support CUDA graphs. | ||
| num_tokens_after_padding = num_tokens | ||
|
|
||
| idx_mapping_list = [ |
There was a problem hiding this comment.
Could we skip updating idx_mapping and recreating sampling metadata if req_ids list hasn't changed?
1 task
|
I'm not sure if this is expected behavior, given that this PR is only a draft, but when testing this PR with If I remove git clone --branch woosuk/model-runner-v2 https://github.com/vllm-project/vllm && cd vllm && git reset --hard 866eef50cae7f9a5f10dbbad8cdf34d07c943f1b
VLLM_USE_PRECOMPILED=1 uv pip install --editable .[flashinfer]
vllm serve RedHatAI/DeepSeek-R1-0528-quantized.w4a16 --tensor-parallel-size 4 -dcp 4 --async-scheduling --served-model-name default --max-model-len 9216Then I send about 50 concurrent requests at it for a minute or so (with several requests sharing a prefix - unsure if that matters), and it crashes. Click to see crash logsNotes:
|
Is it possible you're missing a |
WoosukKwon
force-pushed
the
woosuk/model-runner-v2
branch
from
cf0666e to
09e4b2f
Compare
|
Documentation preview: https://vllm--25266.org.readthedocs.build/en/25266/ |
@josephrocca there are two dcp bugfix !26296 and !27518 (still open) , your case should be fixed by !26296, you can try cherrypick these fixes |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Key Changes
max_model_lenandnum_kv_groupsincreaseModelRunnerOutput&SchedulerOutputTODOs