ViperEkura 816b96a58a test: compare logits_positions rows with assert_close ...

- sliced M=2 and full M=5 lm_head projections may pick different GEMM kernels whose accumulation order differs in the last float32 bits, so torch.equal flakes by machine and thread count

2026-09-05 01:48:06 +08:00

..

config

refactor: generate train and serve CLIs from config-backed option specs

2026-09-03 20:56:11 +08:00

data

test: deduplicate suites and prune low-value cases

2026-09-03 21:54:14 +08:00

extension

refactor: remove bf16 gemm and swiglu kernels and rebuild csrc benchmarks

2026-09-05 01:38:10 +08:00

inference

perf: project only sampled rows through lm_head during prefill

2026-09-05 00:22:15 +08:00

module

test: compare logits_positions rows with assert_close

2026-09-05 01:48:06 +08:00

optim

test: deduplicate suites and prune low-value cases

2026-09-03 21:54:14 +08:00

parallel

style: fix ruff lint warnings

2026-08-07 14:17:48 +08:00

serialization

test: deduplicate suites and prune low-value cases

2026-09-03 21:54:14 +08:00

trainer

test: deduplicate suites and prune low-value cases

2026-09-03 21:54:14 +08:00

conftest.py

perf: pure FP8 fwd/bwd and lean non-transposed GEMM

2026-08-23 15:38:30 +08:00

helpers.py

refactor: deduplicate and restructure test suite

2026-08-20 01:53:28 +08:00