Skip to content

Actions: PrismML-Eng/llama.cpp

Actions

Server

Actions

Loading...
Loading

Show workflow options

Create status badge

Loading
server.yml
203 workflow runs
203 workflow runs

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
speculative: stop plain drafter+target falling back to KV checkpoints Server #258: Pull request #173 synchronize by bri-prism
19m 59s fix/spec-partial-seq-removal
fix/spec-partial-seq-removal
19m 59s
speculative: give the draft context the target's recurrent-state slots Server #257: Pull request #174 opened by bri-prism
20m 33s fix/spec-draft-n-rs-seq
fix/spec-draft-n-rs-seq
20m 33s
speculative: stop plain drafter+target falling back to KV checkpoints Server #256: Pull request #173 opened by bri-prism
12m 16s fix/spec-partial-seq-removal
fix/spec-partial-seq-removal
12m 16s
dflash: AngelSpec DFly drafter support Server #255: Pull request #172 opened by bri-prism
15m 30s feat/dfly-drafter-prism
feat/dfly-drafter-prism
15m 30s
cuda: enable the Hopper wgmma Q1_0/PQ2_0 prefill path by default Server #254: Pull request #171 synchronize by bri-prism
20m 51s perf/hopper-q1-default-on-prism
perf/hopper-q1-default-on-prism
20m 51s
cuda: enable the Hopper wgmma Q1_0/PQ2_0 prefill path by default Server #253: Pull request #171 reopened by bri-prism
4m 27s perf/hopper-q1-default-on-prism
perf/hopper-q1-default-on-prism
4m 27s
cuda: enable the Hopper wgmma Q1_0/PQ2_0 prefill path by default Server #252: Pull request #171 opened by bri-prism
21m 28s perf/hopper-q1-default-on-prism
perf/hopper-q1-default-on-prism
21m 28s
cuda: enable the Hopper wgmma Q1_0/PQ2_0 prefill path by default Server #251: Pull request #169 synchronize by bri-prism
11m 56s perf/hopper-q1-default-on-v7
perf/hopper-q1-default-on-v7
11m 56s
cuda: enable the Hopper wgmma Q1_0/PQ2_0 prefill path by default Server #250: Pull request #169 opened by bri-prism
19m 58s perf/hopper-q1-default-on-v7
perf/hopper-q1-default-on-v7
19m 58s
cuda: keep the PTX L2 prefetch hints off HIP and MUSA Server #249: Pull request #168 opened by khosravipasha
19m 51s fix/hip-prefetch-guard
fix/hip-prefetch-guard
19m 51s
cuda: optimize Bonsai low-bit kernels for GB10 Server #248: Pull request #135 synchronize by khosravipasha
4m 36s pooyakhosravi:perf/gb10-lowbit-kernels
pooyakhosravi:perf/gb10-lowbit-kernels
4m 36s
cuda: block-per-row FWHT for widths 512 and up (+17% / +11% decode on folded bands) Server #247: Pull request #166 synchronize by bri-prism
20m 9s perf/cuda-fwht-block
perf/cuda-fwht-block
20m 9s
cuda: block-per-row FWHT for widths 512 and up (+17% / +11% decode on folded bands) Server #246: Pull request #166 opened by bri-prism
20m 24s perf/cuda-fwht-block
perf/cuda-fwht-block
20m 24s
graph: six fusions for the hybrid-attention decode graph (-25% ops/token; Metal 2B +7.7%, 27B +2.6% decode; CUDA 9B +2.5%) Server #245: Pull request #165 opened by bri-prism
19m 43s perf/metal-fwht-signed
perf/metal-fwht-signed
19m 43s
cuda: run the PTQ1_0 MMQ path at every batch (fixes PTQ1_0 CUDA certification), env var to cap it Server #244: Pull request #164 opened by bri-prism
20m 30s perf/ptq1_0-mmq-cap
perf/ptq1_0-mmq-cap
20m 30s
cuda: query the device type with cudaDeviceGetAttribute, not cudaGetDeviceProperties (+8% decode) Server #243: Pull request #163 opened by bri-prism
20m 29s fix/cuda-dev-type-attr
fix/cuda-dev-type-attr
20m 29s
metal: fix the PTQ1_0 CPY crash and add the missing PTQ1_0 dequant-copy kernels Server #242: Pull request #162 opened by bri-prism
19m 7s fix/metal-cpy-ptq1_0-gate
fix/metal-cpy-ptq1_0-gate
19m 7s
metal: run every folded Hadamard width on the threadgroup FWHT kernel (+4-36% decode) Server #241: Pull request #161 synchronize by bri-prism
20m 38s perf/fwht-2048-wide-prism
perf/fwht-2048-wide-prism
20m 38s
metal: TQ1_0 kernel performance (dense decode +39-55%, prefill +78-90%) Server #240: Pull request #156 synchronize by bri-prism
21m 3s perf/tq1_0-metal-flatten
perf/tq1_0-metal-flatten
21m 3s
metal: run every folded Hadamard width on the threadgroup FWHT kernel (+4-36% decode) Server #239: Pull request #161 opened by bri-prism
19m 45s perf/fwht-2048-wide-prism
perf/fwht-2048-wide-prism
19m 45s
cuda: native PTQ1_0 MMQ tile loader (closes the dequantize-to-cuBLAS prefill fallback) Server #238: Pull request #160 opened by bri-prism
20m 11s perf/ptq1_0-cuda-mmq-prism
perf/ptq1_0-cuda-mmq-prism
20m 11s
metal: TQ1_0 kernel performance (dense decode +39-55%, prefill +78-90%) Server #237: Pull request #156 synchronize by bri-prism
20m 51s perf/tq1_0-metal-flatten
perf/tq1_0-metal-flatten
20m 51s
metal: PTQ1_0 mat-mul dequantize in exact float (+7% prefill) Server #236: Pull request #158 opened by bri-prism
19m 1s perf/ptq1_0-metal-prefill-prism
perf/ptq1_0-metal-prefill-prism
19m 1s
metal: multi-column Q1_0 mat-vec for small verify batches (+ opt-in word-parallel popcount path) Server #235: Pull request #142 synchronize by khosravipasha
20m 18s perf/metal-q1_0-multicol-v7
perf/metal-q1_0-multicol-v7
20m 18s
metal: multi-column Q1_0 mat-vec for small verify batches (+ opt-in word-parallel popcount path) Server #234: Pull request #142 synchronize by khosravipasha
2s perf/metal-q1_0-multicol-v7
perf/metal-q1_0-multicol-v7
2s