COMMITS
August 27, 2026
M
README: quick start first, one command per mode
mhenrichsen committed
M
docker: the entrypoint runs prepare itself; compose up pulls the prebuilt image
mhenrichsen committed
M
README: the prebuilt ghcr image is the default install
mhenrichsen committed
M
CI: build and push the Docker image to ghcr.io on every main push
mhenrichsen committed
August 26, 2026
M
Drafter-free n-gram chains for DFlash2 (VLLM_DFLASH2_CHAIN=1, off by default) (#38)
mhenrichsen committed
M
M
sm80 Marlin repack: grow-only staging for the per-layer transients (#27 candidate)
mhenrichsen committed
M
TP>1: stop pinning the single-card KV_MEM, warn on DFLASH_TOKENS>7 (#40)
mhenrichsen committed
M
int4 per-token-head KV cache with DFlash2: 256k pool on one card (#42)
mhenrichsen committed
August 25, 2026
M
Issue sweep 25.08: flashinfer in Setup, community results, shm ghosts, gotcha 41
mhenrichsen committed
M
Gotcha 9: the resident tower is a cliff config even headless (#36 follow-up)
mhenrichsen committed
D
Offload the vision tower to host RAM, so VISION=1 works with SPEC=dflash2 (#36)
Daniel Donaldson committed
M
Structured output no longer dies at the grammar's end under a speculator (#31)
mhenrichsen committed
August 24, 2026
M
Gotcha 39 is about non-KV headroom, and on WSL2 the shortfall is silent
mhenrichsen committed
August 23, 2026
M
Correct the fp16 guard: the rejection sampler is not bf16-only (#30)
mhenrichsen committed
M
Correct the fp16 guard: the rejection sampler is not bf16-only
mhenrichsen committed
M
MAX_SEQS above 12 at CTX=huge kills the engine (#29)
mhenrichsen committed
M
MAX_SEQS above 12 at CTX=huge kills the engine, and the CG cap does not stop it
mhenrichsen committed
M
Issue sweep: WSL allocator default, bench deps, DFLASH_TOKENS framing, fp16 guard (#28)
mhenrichsen committed
M
Refuse --dtype float16 with a speculator, instead of failing at first attention
mhenrichsen committed
M
DFLASH_TOKENS=15 is a workload setting, not the third of three upgrades
mhenrichsen committed
M
Install pandas, so bench/run_benchmarks.sh runs from a clean install
mhenrichsen committed
M
Turn expandable_segments off on WSL instead of documenting it
mhenrichsen committed
M
Serve images with VISION=1 (#23)
mhenrichsen committed
M
The seats that break CTX=huge, and the workload that does not
mhenrichsen committed
M
The 3.5x for PIECEWISE was the residue bug, not the capture mode
mhenrichsen committed
M
Concurrency is pool-shaped, not seat-shaped
mhenrichsen committed
M
conc_ladder: measure the steady state, the pool and the preemptions
mhenrichsen committed
M
Judge the residue sweeps on how much of the copy came back
mhenrichsen committed
August 22, 2026
M
Refresh the single-stream numbers, measured with the harness that produced them
mhenrichsen committed