intel/auto-round

Pull Requests Commits

test: isolate XPU allocator cache in MoE prefill accuracy tests

Copilot committed 81 days ago

Verified fc018274

refactor: drop MoE prefill over-read scaffolding, use exact [E,K,N] workspace

Copilot committed 81 days ago

Verified 18cd93f2

test: add diagnostic knobs for MoE prefill workspace over-read

Copilot committed 82 days ago

Verified 036ecd72

test: compute MoE prefill reference on CPU

Copilot committed 82 days ago

Verified 2cf1ff1e

fix: zero-initialise MoE test references to avoid stale-memory rows

Copilot committed 82 days ago

Verified 88061bcd

fix: double MoE prefill dequant workspace allocation

Copilot committed 82 days ago

Verified 050d9c87

Revert "fix: pin compat current device to queue's device before default-queue mutation"

Copilot committed 82 days ago

Verified 09ade212

fix: pin compat current device to queue's device before default-queue mutation

Copilot committed 82 days ago

Verified b1304420

fix: synchronise MoE prefill dequant before grouped GEMM reads workspace

Copilot committed 82 days ago

Verified c837bafc

fix: zero-initialise MoE prefill dequant workspace to avoid stale rows

Copilot committed 82 days ago

Verified 232abe03

fix: allocate MoE prefill dequant workspace per call to fix multi-card INT4 races

Copilot committed 82 days ago

Verified 5ebe53c0

fix: correct SageKernelRunner special member function names in sycl_tla_sdpa.hpp

Copilot committed 82 days ago

Verified 342912f6

chore: include allocation sizes in sdpa error messages

Copilot committed 82 days ago

Verified 468ca772

fix: bind sdpa workspace/scratch/sm_count to launch queue device

Copilot committed 82 days ago

Verified a291740b

fix(ark): bind get_stream queue to tensor device for multi-card correctness

Copilot committed 82 days ago

Verified c91d031d

test(ark): isolate ARK_MOE_PREFILL_* env flags per test to fix suite-only failures

Copilot committed 82 days ago

Verified a9109939

fix: route large-M S4 DPAS prefill through validated m_32 geometry

Copilot committed 83 days ago

Verified 5c524593

refactor: extract query_default_queue_sm_count helper in SDPA

Copilot committed 83 days ago

Verified dc21cbaa

fix: bind SDPA grid sizing to the queue's own XPU device

Copilot committed 83 days ago

Verified 864d469b

Merge branch 'main' into copilot/fix-occasional-test-failure

a32543254 committed 83 days ago

Verified 7888dfb5

fix: bind MoE prefill grid sizing to the queue's own XPU device

Copilot committed 83 days ago

Verified 01a44aa5

fix: add triton to requirements for LLMC CPU tests (#2054)

XuehaoSun committed 83 days ago

Verified c774c83b

enhance logging and data loading for AutoScheme and model-free compatibility (#2050)

xin3he committed 84 days ago

Verified 6f45d221

fix: handle router_logits tuple for compatibility with transformers >= 5.13.0 (#2046)

xin3he committed 84 days ago

Verified 4b8468ac

Remove transformers version limit (#2051)

chensuyue committed 84 days ago

Verified 0ea53be4

Fix invalid group reference in revert_checkpoint_conversion_mapping (#2035)

A-S-Manoj committed 84 days ago

Verified 6d1c6db6

Update torch version to 2.13.0 (#2045)

XuehaoSun committed 85 days ago

Verified 32da6a9a

Support auto-round-lib build with oneAPI 2026.0 and work with pytorch xpu 2.13 (#2043)

chensuyue committed 85 days ago

Verified 4d24c405

fix: default int4/int2-sym MoE prefill to bit-exact dequant path (ARK_MOE_PREFILL_DPAS_LOWBIT off)

Copilot committed 87 days ago

Verified be0ef714

Update requirements.txt

a32543254 committed 87 days ago

Verified 2d2c6f49

Older