Repository navigation

All pull requests

Labels (37)Milestones (0)New pull request

tests: portable popcount in arclab_init_hadamard (MSVC build)
#362professorpalmer opened Oct 14, 2026
cuda: exclude sm_75 from the short-query in-place quantized-K/V MMA FA
#361umbertov opened Oct 14, 2026
memory : keep a pending rollback across seq_cp, and default MTP draft depth to 2 on CUDA
#354bri-prism opened Oct 12, 2026Collaborator
1
Merge upstream master into prism (2026-10-11)
#352bri-prism opened Oct 11, 2026Collaborator
8
ggml/meta: tensor split support for Hadamard-folded qwen35 (Bonsai-2)
#351Spacecer2 opened Oct 11, 2026
vulkan: fuse the sign multiply into the Hadamard FWHT
#349elliotmasheal opened Oct 11, 2026
docs: cover the pip SDK route and runtime pitfalls for Windows HIP builds
#344Thatgfsj opened Oct 10, 2026
1
server: --checkpoint-every-nt N, context checkpoints inside a long message (off by default)
#338professorpalmer opened Oct 9, 2026
cuda : take the short-query MMA FA shortcut only for the f16 K/V kernel
#337sb32445 opened Oct 9, 2026
1
grammar: clone finds each stack element's rule by binary search
#336professorpalmer opened Oct 8, 2026
1
json-schema-to-grammar : convert empty anyOf / oneOf / type unions instead of emitting an empty rule
#333professorpalmer opened Oct 7, 2026
cuda: GGML_CUDA_FA_PREFILL_F16=N, f16 prefill for quantized K/V from pool memory (decode stays in place)
#330professorpalmer opened Oct 7, 2026
8
sycl: fuse the PQ2_0/PTQ1_0 XMX mat-muls with their neighbours (depends on #327)
#328kiljoy001 opened Oct 7, 2026
sycl: PQ2_0/PTQ1_0 XMX kernels from TernSYCL, PTQ1_0 kept at 1.75 bpw
#327kiljoy001 opened Oct 7, 2026
sycl: faster concat, get_rows and few-row f16/bf16 mat-vec for short rows
#326kiljoy001 opened Oct 7, 2026
cuda: take the planar PTQ1_0 layout at one column on Turing as well as Ampere
#325professorpalmer opened Oct 7, 2026
cuda: GGML_CUDA_OP_TIMING=1 per-node GPU time breakdown
#323professorpalmer opened Oct 5, 2026
cuda: BATCH_INVARIANT: occupancy-independent FA KV split, PTQ1_0 mat-vec up to 8 columns
#322professorpalmer opened Oct 5, 2026
server: --reasoning-effort-allow/-fallback and --reasoning-max-tokens-floor
#321professorpalmer opened Oct 5, 2026
speculative: --spec-draft-window (MTP drafting at every depth)
#320professorpalmer opened Oct 5, 2026
6
llama/cuda: tiered KV cache (--kv-vram-cells): VRAM head, pinned-host tail in one VMM range
#319professorpalmer opened Oct 5, 2026
llama-quant : refuse PQ2_0/PTQ1_0 output without prism.hadamard metadata
#316cheese-cakee opened Oct 5, 2026
cuda: prefetch the next PTQ1_0 mat-vec's weights into L2 from the last CTAs
#314sb32445 opened Oct 4, 2026
3
cuda: gate (SwiGLU) fused PTQ1_0 mat-vec for 2-4 columns
#310sb32445 opened Oct 4, 2026
cuda: use the transposing concat kernel on all GPUs, not only on GB10
#309sb32445 opened Oct 4, 2026