CUDA: scale q8->f16 cache reserve on >=112 GiB cards (fixes session OOM on large models) - #472
Open
slackarea wants to merge 1 commit into
Open
CUDA: scale q8->f16 cache reserve on >=112 GiB cards (fixes session OOM on large models)#472slackarea wants to merge 1 commit into
slackarea wants to merge 1 commit into
Conversation
cuda_q8_f16_cache_reserve_bytes() returned a flat 512 MiB reserve once total VRAM >= 112 GiB, instead of the 5% / 4 GiB-min rule used below that. The q8->f16 dequant cache is eager and fills HBM down to the reserve, so on a large model the session/context graph allocated after model load OOMs at session creation even though the weights themselves fit. WEIGHT_CACHE_LIMIT_GB does not bound this cache, and loading an MTP model disables it and hides the issue. Drop the >=112 GiB special case so every card uses 5% / 4 GiB-min. This is the CUDA twin of antirez#446 (same bug on the ROCm runtime, q4q2). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQVgY7rXrksjtBjPFSCnMH
fulvius31
added a commit
to fulvius31/ds4
that referenced
this pull request
Jul 24, 2026
…ez#524) server: when thinking mode is on and the model never emits </think> (typically truncated at max_tokens mid-thought), surface the whole text as unfinished reasoning with empty content instead of letting it leak into content — applied to both the DSML and GLM tool-call parse paths (upstream antirez#524, adapted to the split parser). Verified live: /v1/messages now returns a thinking block plus empty text for the truncated case; closed-thinking responses are unchanged. cuda: treat every model range as accessible when the HMM direct path is active (upstream antirez#158, the is_cached hunk only). The PR's full-model prefetch half is deliberately not ported: prefetching a model larger than RAM would thrash the page cache on the SSD-streaming Spark boxes. Dormant in current configs; single-token identity exact. Reviewed against the queue: antirez#497 already fixed here (b715001), antirez#472 superseded by the newer bounded-reserve design, antirez#513 not reachable (expert tiles ship disabled for iq2-down), antirez#460/antirez#528/antirez#504 deferred as future work. Co-Authored-By: Claude Fable 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
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.
cuda_q8_f16_cache_reserve_bytes()returns a flat 512 MiB reserve once total VRAM>= 112 GiB, instead of the5% / 4 GiB-minrule used below that. The q8->f16 dequant cache is eager and fills HBM down to the reserve, so on a large model the session/context graph allocated after model load OOMs at session creation even though the weights themselves fit.DS4_CUDA_WEIGHT_CACHE_LIMIT_GBdoes not bound this cache, and loading an MTP model disables it and hides the issue.Fix: drop the
>= 112 GiBspecial case so every card uses5% / 4 GiB-min. This is the CUDA twin of #446 (same bug on the ROCm runtime, q4q2).Evidence (2x H200 NVL, 143 GB, no NVLink, CUDA 13.2)
A ~252 GB model split across 2 GPUs (
--role coordinator --layers 0:36+--role worker --layers 37:output), no SSD streaming:CUDA tensor alloc failed: out of memoryat session create —cached=14.37 GiB free=0.54 GiB reserve=0.50 GiB.cached=7.87 GiB free=7.04 GiB reserve=6.99 GiB, session creates, runs fully resident — prefill 39 t/s, generation 16 t/s.Non-regression (DeepSeek-V4 Flash, single H200, IQ2XXS)
Same build: prefill 97.0 t/s, generation 40.8 t/s (matches baseline). The change only enlarges the optional cache's reserve on
>= 112 GiBcards; correctness is unaffected (logits identical — the cache is a dequant acceleration path), and on a single card a small model leaves plenty of free VRAM so the reserve does not bind.🤖 Generated with Claude Code