[Doc] Record Qwen3.8 prefill output-quality gate - #403
Merged
yangzhuxinyzx merged 26 commits intoAug 29, 2026
Merged
yangzhuxinyzx merged 26 commits into
yangzhuxinyzx merged 26 commits into
Conversation
Use a fresh async H2D tensor for each block-ID batch so a later scheduler step cannot overwrite pinned storage while an earlier copy is still in flight. Add a CUDA-stream regression test that holds both copies behind pending GPU work. Upstream-Reference: vllm-project/vllm#48399 Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Treat each index weight_map assignment as authoritative within its shard so reused safetensors files cannot inject stale or incompatible tensors. Assisted-by: OpenAI Codex Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…t paths Mirror of the review fix applied to the upstream submission (vllm-project/vllm#54230): the enable_multithread_load branch bypassed the per-file index allowlist, so reused shards kept yielding unindexed and duplicate tensors with nondeterministic overwrite order under as_completed. Pass indexed_weights_by_file into multi_thread_safetensors_weights_iterator and filter inside each worker. fastsafetensors/instanttensor iterators cannot filter within a shard, so indexed-subset checkpoints now fall back to the filtered standard iterator with a warning. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Assisted-by: OpenAI Codex Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…promotion-20260829-042116 [Core][SM70] Promote DFlash2 NVFP4 prefill route
Add exact-topology checkpoint-FP16 row GEMV, fused GDN input and HyperConnection routes, and shorten exact QSA top-k to a score-pivot compaction. The matched TP4 no-MTP I8192/O512 gate reaches 80.732 tok/s from 65.864 tok/s while the frozen GSM8K and long-output health gates pass. Assisted-by: OpenAI Codex Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…lter [Bugfix] Honor safetensors index tensor assignments
[Bugfix] Avoid KVBlockZeroer H2D staging races
…20260829-030630 [Kernel][SM70] Accelerate exact Qwen3.8 decode
…lease-audit-20260829-052327 Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com> # Conflicts: # docs/design/sm70_v100_migration_control.md
Make the incompatible TokenSpeed MLA package optional and copy tracked FlashQLA sources into the csrc wheel stage.\n\nReported-by: ga-it <42608074+ga-it@users.noreply.github.com> Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…20260829-052327 [Core][SM70] Preserve DFlash2 decode across NVFP4 prefill
…-build-closure-20260829-074938
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…ard-20260829-080317 [Bugfix][SM70] Gate DFlash2 verifier by native capability
…re-20260829-074938 [CI/Build][SM70] Close wheel build inputs
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@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
2 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.
Purpose
Record the matched output-quality audit for the Qwen3.8 Flash Next NVFP4 TP4 prefill optimization merged in #393. This PR changes documentation only; runtime behavior is unchanged.
Test Plan
Test Result
git diff --check, typos, markdownlint, and all applicable pre-commit hooks pass.