FP8 KV-Cache Compatibility for Hybrid vLLM Models#1899
Open
YJHMITWEB wants to merge 10 commits into
Open
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces support for serialized blockwise FP8 weight synchronization and persistent FP8 parameter training, alongside memory optimization features like CPU-resident microbatches and a colocated worker memory barrier with hard eviction. It also adds a compatibility patch for vLLM FP8 KV-cache sleep/wake resets on nested cache entries. The feedback highlights an optimization opportunity in the weight extraction process, suggesting that blockwise FP8 quantization should only be performed on rank 0 to avoid redundant GPU/CPU compute and memory usage on other ranks.
YJHMITWEB
force-pushed
the
fp8-rl-kv-cache
branch
4 times, most recently
from
July 16, 2026 18:39
fa82f40 to
8e47ee8
Compare
YJHMITWEB
force-pushed
the
fp8-rl-kv-cache
branch
from
July 17, 2026 00:26
8e47ee8 to
5f8791a
Compare
PPORayActorGroup stored `pg` and `num_gpus_per_actor` on self but nothing ever read them back (the constructor already passes both directly to _initiate_actors). Drop the two dead assignments; the constructor params are unchanged.
Make the previously opt-in `cpu_resident_microbatches` behavior the only path and remove the config flag. Each microbatch is now moved to the device just before its forward step (in MegatronModelWrapper.forward_step), instead of moving the whole mini-batch up front. This caps resident input-tensor memory at one microbatch and is a no-op when a tensor is already on device. The per-microbatch transfer is the single choke point for every forward and forward-backward pass, so the up-front `.to(device)` calls in the worker are removed; nothing downstream reads the mini-batch tensors on-device after the microbatch loop (only chunk()/BatchIterator, which slice on CPU, and metadata). Removes the MegatronConfig field, the wrapper constructor param, the worker plumbing, and the CLI-override test for the flag.
YJHMITWEB
force-pushed
the
fp8-rl-kv-cache
branch
from
July 21, 2026 19:46
5f8791a to
0003b26
Compare
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Summary
This PR enables vLLM's existing
fp8_e4m3KV-cache mode for hybrid attention/Mamba models that use nested cache structures. It does not change weight quantization, parameter storage, or weight syncChanges
Quality
All FP8 runs show perfect alignment with the BF16 baseline.
Qwen3.5-4B runs: https://wandb.ai/sky-posttraining-uc-berkeley/qwen35_4b_dapo_mainclean_fp8align_fastbatch_bf16_fp8_20260630?nw=nwuserjinghanyao1
Qwen3.5-9B runs: https://wandb.ai/sky-posttraining-uc-berkeley/qwen35_9b_h100_tp2_bf16_fp8_prready_20260709?nw=nwuserjinghanyao1
Qwen3.5-35B-A3B runs: https://wandb.ai/sky-posttraining-uc-berkeley/qwen35_35b_a3b_fp8_amaxeps_20260705?nw=nwuserjinghanyao1