Conversation
Integrate the Qwen3.8 Flash Next model and PLE CPU offload work from vLLM PRs #53896 and #53899, then carry the validated 1Cat SM70 compatibility layer for FP16 activations, hybrid caches, NVFP4 emulation, and top-token sampling. The included TP4 smoke covers checkpoint loading, PLE offload, prompt prefill, and one autoregressive decode step on four V100 32GB GPUs. Co-authored-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com> Co-authored-by: zjy0516 <riverclouds.zhu@qq.com> Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Allow grouped AWQ scale and zero-point tensors to expand exactly when a tensor-parallel expert shard is narrower than the checkpoint group size. Add unit coverage and a Qwen3.8 Flash Next TP4 loader fixture. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Record deterministic TP4 prefill and decode samples for the NVFP4 emulation path, including per-case warmups, incremental JSON output, and a low-headroom KV cache control. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
4 tasks done
Match the 1Cat mRoPE, FP8 skip, short-conv metadata, and host-to-device helper signatures. Narrow optional FP8 test scales so the upstream pre-commit mypy matrix can validate the imported code. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Widen the 1Cat hybrid state-shape protocol to the current vLLM shape union and document intentional Qwen4Exp overrides for the older 1Cat Qwen3Next and pipeline-parallel interfaces. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Apply the pinned Ruff 0.14.0 formatter to the two PR-specific files that differ from the current 1Cat main formatting baseline. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
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
1 participant
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.
Scope
Local 1Cat integration for Qwen3.8-Flash-Next on 4x V100. This does not duplicate upstream vLLM model work: it adapts the still-open vllm-project/vllm#53896 and PLE offload work in vllm-project/vllm#53899 to the 1Cat SM70 runtime.
Changes
Validation
Known limitations
NVFP4 on V100 uses on-the-fly emulation and is not a production recommendation. The measured path is eager, single-stream, max length 256, 64 MiB KV per GPU, prefix caching off, async scheduling off, and MTP off. The intended production checkpoint is activation-aware AWQ W4A16 g32 with PLE on CPU.
AI assistance
OpenAI Codex assisted with implementation, tests, runtime diagnostics, and benchmark tooling. This PR remains Draft and requires human review before merge.