Pin flip to vLLM 0.29.0: the series re-exported at fuzz 0, KVarN on 0.29, the #114 moves, a cold-boot profiling fix, and 0.28 vs 0.29 on every run setting (#106 part two) - #148

Merged

mhenrichsen merged 23 commits into

syv-ai:mainfrom

cpuchip:port-0.29-hq

Sep 23, 2026

Commits on Sep 18, 2026

port 0.29: pin vllm==0.29.0; the series re-exported from cpuchip/vllm qwen38/0.29 at 337efb79f (35 patches, 0 fuzz, replay identical to the fork tip); KVarN 0.29 patches and modules; retirements li…
cpuchip
committed
c0ced23
port 0.29: launcher, verify, Dockerfile comment, PR-triggered image build, docs (carried from port-0.29-bae2023 onto the split README)
cpuchip
committed
7e6d9f4
#114: register the spec-decode-attn knobs in the patch that reads them
cpuchip
committed
7a6b00f
0.29: memory-profile-after-warmup, so a cold compile cache does not inflate the profiled transient peak and refuse the KV cache the warm boot grants (cold first boots failed on B and E on a WSL2 40…
cpuchip
committed
1780966
docs/vllm-0.29.md: the 0.28 vs 0.29 campaign on every run setting (WSL2 4090), and the cold-boot falsifier on both boxes (default profile fixed; CTX=huge cold boot still refused on the WSL2 4090, o…
cpuchip
committed
74ce361
PATCHES.md: name both fork branches, and the four series entries that had no row
cpuchip
committed
a2de01e
PATCHES.md: row for memory-profile-after-warmup, and the -hq branch moved to bf29fa109
cpuchip
committed
fdc8589
docs: re-point the version-specific claims the pin moved under
cpuchip
committed
4cb70cb
docs/vllm-0.29.md: the native 3090 CTX=huge cold boot with the fix passes; the open item narrows to the WSL2 4090's first boot
cpuchip
committed
8b4316e
memory-profile-after-warmup: header corrected (0.28.0 also refuses the cold default boot on the native 3090; the WSL2 4090 CTX=huge residual is the graph estimate), same hunks, fork tip 512a9699c; …
cpuchip
committed
eac21d0
Merge syv-ai main (field tests, #146) into port-0.29-hq; README: the 0.29 row of the field-tests table now names this branch's pin
cpuchip
committed
82b9441
docs/vllm-0.29.md: VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 as the CTX=huge path on the WSL2 4090 (cold 4.47 GiB, warm 4.68, actual graph pool 0.0), the knob in the compile-cache key, the equival…
cpuchip
committed
9ccaf97
docs/vllm-0.29.md: the estimate-off path is verified only where the actual graph pool reads 0.0 GiB; the native 3090 reads 0.11 to 1.87 GiB by profile, so it under-reserves there
cpuchip
committed
cc894e9
cudagraph-memory-from-allocator: measure the captured graph pool by the allocator's reserved bytes (WSL2's driver reading is pinned at zero after KV allocation and collapses by 5.44 GiB during a co…
cpuchip
committed
f07554d
docs/vllm-0.29.md: the native 3090 twelve-boot A/B (three settings with both arms, three 0.28 refusals with numbers), the measured driver-to-allocator residue (0.02 and 0.03 GiB), the KV the alloca…
cpuchip
committed
67858f0

Navigation Menu

Uh oh!

Pin flip to vLLM 0.29.0: the series re-exported at fuzz 0, KVarN on 0.29, the #114 moves, a cold-boot profiling fix, and 0.28 vs 0.29 on every run setting (#106 part two) - #148

Pin flip to vLLM 0.29.0: the series re-exported at fuzz 0, KVarN on 0.29, the #114 moves, a cold-boot profiling fix, and 0.28 vs 0.29 on every run setting (#106 part two)#148
mhenrichsen merged 23 commits into
syv-ai:mainfrom
cpuchip:port-0.29-hq

Commits on Sep 18, 2026

port 0.29: pin vllm==0.29.0; the series re-exported from cpuchip/vllm qwen38/0.29 at 337efb79f (35 patches, 0 fuzz, replay identical to the fork tip); KVarN 0.29 patches and modules; retirements li…

port 0.29: launcher, verify, Dockerfile comment, PR-triggered image build, docs (carried from port-0.29-bae2023 onto the split README)

#114: register the spec-decode-attn knobs in the patch that reads them

0.29: memory-profile-after-warmup, so a cold compile cache does not inflate the profiled transient peak and refuse the KV cache the warm boot grants (cold first boots failed on B and E on a WSL2 40…

docs/vllm-0.29.md: the 0.28 vs 0.29 campaign on every run setting (WSL2 4090), and the cold-boot falsifier on both boxes (default profile fixed; CTX=huge cold boot still refused on the WSL2 4090, o…

PATCHES.md: name both fork branches, and the four series entries that had no row

PATCHES.md: row for memory-profile-after-warmup, and the -hq branch moved to bf29fa109

docs: re-point the version-specific claims the pin moved under

docs/vllm-0.29.md: the native 3090 CTX=huge cold boot with the fix passes; the open item narrows to the WSL2 4090's first boot

memory-profile-after-warmup: header corrected (0.28.0 also refuses the cold default boot on the native 3090; the WSL2 4090 CTX=huge residual is the graph estimate), same hunks, fork tip 512a9699c; …

Merge syv-ai main (field tests, #146) into port-0.29-hq; README: the 0.29 row of the field-tests table now names this branch's pin

docs/vllm-0.29.md: VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 as the CTX=huge path on the WSL2 4090 (cold 4.47 GiB, warm 4.68, actual graph pool 0.0), the knob in the compile-cache key, the equival…

docs/vllm-0.29.md: the estimate-off path is verified only where the actual graph pool reads 0.0 GiB; the native 3090 reads 0.11 to 1.87 GiB by profile, so it under-reserves there

cudagraph-memory-from-allocator: measure the captured graph pool by the allocator's reserved bytes (WSL2's driver reading is pinned at zero after KV allocation and collapses by 5.44 GiB during a co…

docs/vllm-0.29.md: the native 3090 twelve-boot A/B (three settings with both arms, three 0.28 refusals with numbers), the measured driver-to-allocator residue (0.02 and 0.03 GiB), the KV the alloca…

Commits on Sep 22, 2026

Merge syv-ai main (13b30ea) into port-0.29-hq; the 0.29 series carries #142 and the four new patches

triton-spec-attn-fp8-kv: re-export without the copy of its own diff that its fork commit message carried

docs/vllm-0.29.md: the native 3090 row names the box, not a seat; PATCHES.md: dflash2-backport is the one file with no export point, and qwen38/0.29-hq2 is a rewrite of -hq from hybrid-sw-block-pro…

Merge syv-ai main 13008f6 (#179, #180) into port-0.29-hq

single-user: on 0.29, #179's retention interval goes in as --prefix-cache-retention-interval

single-user: the retention comment gives 0.29's per-turn reuse (80.0% on the first reuse after a cold turn, then 99.3-99.4%, against 0% dense)

Commits on Sep 23, 2026

Merge syv-ai main (ee15878: #172, #181) into port-0.29-hq

batch: GPU_UTIL defaults to 0.95 on 0.29 (0.972 OOMs in warmup, #182)