Repository navigation

All pull requests

colibri (Mistral-Large-4 vision) serve: a picture announced for a refused request does not go to the next one
#2771KairaCode opened Nov 11, 20263 of 4 ta 8000 sks
qwen38 prefix store: with the MTP head on, a take leaves in place only the blocks below the row the slot's next token made
#2770KairaCode opened Nov 11, 2026
tests(qwen38): read the reuse source in test_qwen38_prefix_serve without the CR that Windows' stderr adds
#2769KairaCode opened Nov 11, 20262 of 3 tasks
qwen38 serve: a picture's rows go with a request refused with CONTEXT_EXCEEDED
#2768KairaCode opened Nov 11, 2026
deepseek-v4 CUDA: round the input of the generic 128-block fp8 matvecs to E4M3 per 128 on the device, as the CPU reference and the fused kernels do
#2767KairaCode opened Nov 11, 20263 of 4 tasks
prefix store: refuse a save its budget cannot hold before evicting, take back a failed save's blocks, hold the RAM floor across a whole save, keep a chain's system-prompt checkpoint
#2766KairaCode opened Nov 11, 20262 of 3 tasks
Qwen3.8-27B speculative decoding on CUDA: verifies on the tensor cores for every GGUF type, DFlash2, drafts in lanes, v3.1 MTP files
#2765JustVugg opened Nov 11, 2026Owner
deepseek-v4: upload the exact CUDA head even when the RAM plan streams it (read it for the upload, then free it)
#2763KairaCode opened Nov 11, 20263 of 4 tasks
qwen38 serve: under COLI_TIMERS=1 report each request's prefill and decode banks and where its experts came from
#2762KairaCode opened Nov 11, 2026
qwen38/qwen36: print 'Text :' on its own line after the reference 'C engine :' ids
#2761KairaCode opened Nov 11, 2026
deepseek-v4: CUDA_ARCH=portable-pre-ampere needs CUDA <= 12.9; say so where the preset is recommended
#2760KairaCode opened Nov 11, 2026
fix(datapoint): reject non-finite memory budgets
#2756Stamina9 opened Nov 11, 2026Contributor
build: cuda-dsv4-dll on Windows accepts CUDA_ARCH=portable-pre-ampere and NO_TC=1
#2755KairaCode opened Nov 11, 2026
tests: let the int2/int4 converter tests delete their temp dirs on Windows
#2754KairaCode opened Nov 11, 2026
deepseek-v4: load an absolute COLI_DSV4_DLL path as given; keep the CWD out of the CUDA runtime search
#2752KairaCode opened Nov 10, 2026
tests(qwen38): keep compat.h's setenv on Windows so the qwen38 tests build with MinGW
#2751KairaCode opened Nov 10, 2026
Add hackernews badge
#2749murtazayusuf opened Nov 10, 20266 tasks
qwen38: free a placed trunk matrix's BF16 rows as soon as its int8 copy exists, not after the whole trunk
#2748KairaCode opened Nov 10, 2026
1
qwen38 CUDA chain: a device-memory refusal gives the K/V mirror back instead of holding it for the process
#2747KairaCode opened Nov 10, 2026
qwen38: keep the CUDA chain's device memory out of the FP8 expert tier's budget, as the GGUF arena already does
#2746KairaCode opened Nov 10, 2026
deepseek-v4: pack DeepGEMM scales only for shapes the DeepGEMM kernels take (no zero grid, no divide by zero)
#2745KairaCode opened Nov 10, 2026
build: ignore and clean c/deepseek_v4.ldflags, the V4 link stamp from #2652
#2744KairaCode opened Nov 10, 2026
web: connect to the address the public check passed
#2742divya0795 opened Nov 10, 2026
docs: building the CUDA DLL from a conda-layout CUDA tree on Windows
#2740tylnexttime opened Nov 10, 20262 of 5 tasks
cuda: build without nvcc #221 on Windows and an unused ANS-only global
#2739tylnexttime opened Nov 10, 20263 of 5 tasks

Pull requests