Repository navigation
feat(vllm): bump to v0.32.0 — KV 429 319 (+1.1%), full-window 400 eliminated - #122
Merged
Merged
Conversation
…minated GO user 19/10 (« Go pour le bump maintenant »). Fenêtre : down ~20:13Z, StartedAt 20:18:09Z, healthy ~20:24Z (~11 min, boot froid sans stall). Gates 11/11 effectifs, FORCING all-200, zéro erreur, config byte-identique. KV 429 319 vs 424 610 (+1,1 %) à config égale. Bench post-boot 928,4 t/s @n=48 (≥ meilleure référence historique) — le delta apparent vs le baseline du jour n'est PAS retenu (téléchargement d'image concurrent + moteur âgé, cf. NOTES). vllm-project#57035 vérifié en direct : prompt 260 012 + max_tokens 4096 > 262 144 → HTTP 200 (400 en v0.31) — la classe du 400 vllm-project#431 est éliminée côté moteur. Nouveaux harnais réutilisables : gates_v0320.py, window_full_test.py. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fenêtre (GO user 19/10)
Down ~20:13Z → StartedAt 20:18:09Z → healthy ~20:24Z (~11 min, boot froid sur volume
…-v0320, aucun stall). Rollback armé : 3 lignes + volumes v0310/v1 intacts.Résultats
vLLM 0.32.0confirmé dans le conteneur,Using V2 Model Runner, args byte-identiques (diff = 3 lignes de profil).max_tokens=8+ thinking par défaut), FORCING (forwards (4096, 8192]) all-200, zéro signature d'erreur.max_tokens=4096(264 108 > 262 144) → HTTP 200 (v0.31 : 400, mesuré 09/10). La classe du 400 de RuntimeError: Failed to import transformers.models.bloom.modeling_bloom because of the following error (look up to see its traceback): /usr/local/lib/python3.8/dist-packages/transformer_engine_extensions.cpython-38-x86_64-linux-gnu.so: undefined symbol: _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_RKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE vllm-project/vllm#431 est éliminée côté moteur ; reste licite un prompt seul > 262 144.Contenu
medium-swift15-27b.yml: 3 lignes (image, mount compile-cache, déclaration volume)._iteration_log/v0320_bump/: NOTES complet + harnais réutilisablesgates_v0320.py(13 gates, PNG vision stdlib) etwindow_full_test.py(test de réservation calibré par/tokenize).Suite
Phase 2 (GO user) : éval NVFP4-KV sur SM89 — la piste KV ouverte par la v0.32 (vllm-project#46963).
🤖 Generated with Claude Code