intel/auto-round

Pull Requests Commits

feat: add W4A8 (int4 weight / int8 compute) ARK XPU MoE kernel

Copilot committed 10 days ago

Verified ef97754e

Merge branch 'copilot/optimize-int4-moe-performance' into W4A8 MoE work

Copilot committed 10 days ago

Verified 05ad2e16

docs: document Qwen3-MoE int4-sym prefill/decode perf tests

Copilot committed 10 days ago

Verified 42949b74

test: add qwen3-moe int4-sym prefill and decode perf benchmarks

Copilot committed 10 days ago

Verified 353543da

feat: add per-head fp8 kv attention calibration (#2044)

zeel2104 committed 11 days ago

Verified f17d9cd4

fix performance regression bug (#2136)

n1ck-guo committed 11 days ago

Verified f8ad63f4

Enhance unit tests with timeout handling and stability improvements (#2135)

XuehaoSun committed 12 days ago

Verified e4a02ccd

perf: raise the MoE auto-dispatch decode cutoff from 32 to 128 tokens

Copilot committed 14 days ago

Verified 03ca8774

perf: N-block the FP8 K-split decode GEMV and drop the per-call routing sync

Copilot committed 14 days ago

Verified 51328f36

Add python 3.14t compatibility tests (#2110)

XuehaoSun committed 14 days ago

Verified 98f2cdb9

support one more mxfp4 variant (#2132)

wenhuach21 committed 14 days ago

Verified f706da67

Add llmc test back (#2131)

yiliu30 committed 14 days ago

Verified 1f6b2151

docs: describe the K-split FP8 decode GEMV (EN + CN)

Copilot committed 14 days ago

Verified e43b98f7

perf: add K-split lane-parallel FP8 MoE decode GEMV

Copilot committed 14 days ago

Verified 7102f0ec

fix: restore test_perf_fp8_per_tensor definition; make fp8_decode_mode allocation-free

Copilot committed 14 days ago

Verified c6f67a13

test: FP8 decode-mode + DPAS parity/perf coverage; docs: FP8 decode section (EN+CN)

Copilot committed 14 days ago

Verified 3423da8f

perf: word-native FP8 decode + decode-tuned FP8 DPAS dispatch

Copilot committed 14 days ago

Verified 6cd81921

docs: update int4 decode perf notes for word-native nibble decode

Copilot committed 14 days ago

Verified a2827941

perf: speed up int4-sym MoE decode with word-native nibble decode

Copilot committed 14 days ago

Verified 610a005d

test,docs: cover the new int4 decode paths and document the tuning knobs

Copilot committed 15 days ago

Verified 4ec9d7fe

perf(ark): vectorize int4 decode GEMV loads, hoist activation sums, pool the repack buffer

Copilot committed 15 days ago

Verified e4d25c5d

perf(ark): decode int4-sym nibbles through the asym unsigned path (^0x88 sign-flip)

Copilot committed 15 days ago

Verified 211e95f8

fix sparge on torch2.13 (#2130)

yiliu30 committed 15 days ago

Verified d8a30059

perf(ark): route decode-sized int4-sym moe through the asym scalar GEMV

Copilot committed 15 days ago

Verified 1f7b6313

perf(ark): use A_avg_M tile ladder for int4 moe sym decode (match vllm-xpu-kernels)

Copilot committed 15 days ago

Verified 5b88c2ec

Refator UT and improve code coverage (#2057)

lvliang-intel committed 15 days ago

Verified 7a75218a

perf: trim device-timeline overhead on S4 DPAS decode hot path

Copilot committed 15 days ago

Verified ba2c18d7

perf: reuse persistent atomic buffer in S4 DPAS dispatch

Copilot committed 15 days ago

Verified 33960828

docs+test: document decode S4 DPAS path and add A/B parity assertion

Copilot committed 15 days ago

Verified a3c82ced

feat: add dedicated w4a16_policy_m_8 decode DPAS dispatch for S4 sym

Copilot committed 15 days ago

Verified c8ef1b8e

Newer Older