Skip to content

Actions: gittensor-ai-lab/sparkinfer

Actions

build-attested-binaries

Actions

Loading...
Loading

Show workflow options

Create status badge

Loading
build-attested-binaries.yml
1,276 workflow runs
1,276 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
perf(bonsai): the decode shadow reads the attention and output projec… build-attested-binaries #1332: Commit 7212971 pushed by skyrocket2026
In progress main
main
In progress
perf(kernels): the fused Q4_K prefill GEMM decodes on its own warps a… build-attested-binaries #1331: Commit 2fa6fd5 pushed by skyrocket2026
12m 53s main
main
12m 53s
perf(bonsai): the decode shadow reads the attention and output projections in their ternary blocks too, with the GDN norm and the attention gate folded into their rotation (1.11x cb-decode @c16 on Ternary-Bonsai-2) build-attested-binaries #1330: Pull request #1166 opened by FranDev132
8m 31s FranDev132:perf/bonsai-shadow-attn-out
FranDev132:perf/bonsai-shadow-attn-out
8m 31s
perf(kernels): the fused Q4_K prefill GEMM decodes on its own warps and takes a 256-row tile past 256 rows (1.28x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1329: Pull request #1163 synchronize by FranDev132
14m 29s FranDev132:perf/bonsai-prefill-qb-ws-tall
FranDev132:perf/bonsai-prefill-qb-ws-tall
14m 29s
perf(bonsai): packed rows read the decode shadow's FFN on the int8 te… build-attested-binaries #1328: Commit 61ccf92 pushed by skyrocket2026
16m 17s main
main
16m 17s
perf(kernels): the fused Q4_K prefill GEMM decodes on its own warps and takes a 256-row tile past 256 rows (1.28x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1327: Pull request #1163 synchronize by FranDev132
13m 0s FranDev132:perf/bonsai-prefill-qb-ws-tall
FranDev132:perf/bonsai-prefill-qb-ws-tall
13m 0s
perf(kernels): the fused Q4_K prefill GEMM decodes on its own warps and takes a 256-row tile past 256 rows (1.28x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1326: Pull request #1163 synchronize by FranDev132
10m 3s FranDev132:perf/bonsai-prefill-qb-ws-tall
FranDev132:perf/bonsai-prefill-qb-ws-tall
10m 3s
perf(bonsai): packed rows read the decode shadow's FFN on the int8 tensor cores, one weight read for every row (1.33x cb-decode @c8 on Ternary-Bonsai-2) build-attested-binaries #1325: Pull request #1164 synchronize by FranDev132
12m 11s FranDev132:perf/bonsai-shadow-ternary-tc
FranDev132:perf/bonsai-shadow-ternary-tc
12m 11s
perf(bonsai): gate/up stay in their ternary blocks, and packed batche… build-attested-binaries #1324: Commit dc723a4 pushed by skyrocket2026
19m 30s main
main
19m 30s
perf(bonsai): packed rows read the decode shadow's FFN on the int8 tensor cores, one weight read for every row (1.33x cb-decode @c8 on Ternary-Bonsai-2) build-attested-binaries #1323: Pull request #1164 opened by FranDev132
12m 24s FranDev132:perf/bonsai-shadow-ternary-tc
FranDev132:perf/bonsai-shadow-ternary-tc
12m 24s
perf(bonsai): gate/up stay in their ternary blocks, and packed batches past 8 rows read the whole FFN there on the int8 tensor cores (1.43x cb-decode @c16 on Ternary-Bonsai-2) build-attested-binaries #1322: Pull request #1161 synchronize by FranDev132
11m 37s FranDev132:perf/bonsai-native-gate-up
FranDev132:perf/bonsai-native-gate-up
11m 37s
perf(kernels): the fused Q4_K prefill GEMM decodes on its own warps and takes a 256-row tile past 256 rows (1.28x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1321: Pull request #1163 opened by FranDev132
10m 36s FranDev132:perf/bonsai-prefill-qb-ws-tall
FranDev132:perf/bonsai-prefill-qb-ws-tall
10m 36s
perf(kernels): the long-context int8 prefill GEMM loses its bank conf… build-attested-binaries #1320: Commit 20a4fbd pushed by skyrocket2026
13m 13s main
main
13m 13s
fix(bonsai): release the decode shadow's VRAM for batched prefill, no… build-attested-binaries #1319: Commit cc6cfc0 pushed by skyrocket2026
12m 32s main
main
12m 32s
perf(kernels): the long-context int8 prefill GEMM loses its bank conflicts, takes a 64x64 warp tile and forms the FFN's SwiGLU in its epilogue (1.07x prefill @4k on Ternary-Bonsai-2) build-attested-binaries #1318: Pull request #1162 opened by FranDev132
9m 3s FranDev132:perf/dense-prefill-i8-gemm-w4
FranDev132:perf/dense-prefill-i8-gemm-w4
9m 3s
refactor(qwen35_prefill): improve GDN handling and context management… build-attested-binaries #1317: Commit 053ef27 pushed by skyrocket2026
12m 46s main
main
12m 46s
fix(bonsai): release the decode shadow's VRAM for batched prefill, not only a new session build-attested-binaries #1316: Pull request #1160 reopened by skyrocket2026
12m 34s inference2026:fix/bonsai-decode-shadow-vram-release
inference2026:fix/bonsai-decode-shadow-vram-release
12m 34s
perf(bonsai): gate/up stay in their ternary blocks, and packed batches past 8 rows read the whole FFN there on the int8 tensor cores (1.43x cb-decode @c16 on Ternary-Bonsai-2) build-attested-binaries #1315: Pull request #1161 opened by FranDev132
10m 8s FranDev132:perf/bonsai-native-gate-up
FranDev132:perf/bonsai-native-gate-up
10m 8s
fix(bonsai): release the decode shadow's VRAM for batched prefill, not only a new session build-attested-binaries #1314: Pull request #1160 opened by inference2026
12m 38s inference2026:fix/bonsai-decode-shadow-vram-release
inference2026:fix/bonsai-decode-shadow-vram-release
12m 38s
perf(kernels): pack bf16 V into the wmma tile for dense-GGUF full attention (1.013x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1313: Pull request #1159 opened by JamesR111
10m 24s JamesR111:perf/bonsai-prefill-attn-vpack
JamesR111:perf/bonsai-prefill-attn-vpack
10m 24s
perf(qwen35): fill the device on the N=512 GDN bf16 GEMMs and fuse th… build-attested-binaries #1312: Commit c700f98 pushed by skyrocket2026
13m 56s main
main
13m 56s
perf(bonsai): decode reads the head and GDN qkv/z from ternary copies through a tensor-core GEMV (1.07x decode @128, 1.08x cb-decode @c4 on Ternary-Bonsai-2) build-attested-binaries #1311: Pull request #1157 opened by widecloud
11m 39s widecloud:perf-bonsai-ternary-decode
widecloud:perf-bonsai-ternary-decode
11m 39s
perf(bonsai): decode and small packed batches read the FFN from a ternary copy through a dp4a GEMV (1.35x decode @128) build-attested-binaries #1310: Pull request #1154 synchronize by kaivaryn
13m 5s kaivaryn:perf/bonsai-ffn-shadow-v2
kaivaryn:perf/bonsai-ffn-shadow-v2
13m 5s
perf(qwen35): dense-GGUF GDN prefill at 512 takes fused quantized-B, not bf16 (1.29x prefill @512 on Ternary-Bonsai-2) build-attested-binaries #1309: Pull request #1156 opened by DripMicro
12m 7s DripMicro:perf/bonsai-i8-gdn-512
DripMicro:perf/bonsai-i8-gdn-512
12m 7s
perf(kernels): dense Q4_K cb-decode takes the tensor-core down and in… build-attested-binaries #1308: Commit f90d6b5 pushed by skyrocket2026
12m 56s main
main
12m 56s