yangzhuxinyzx ce0ad19520 [Test] Skip CUDA attention regressions when no device is visible ...

Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

2026-10-02 22:54:32 +08:00

..

attention

[Test] Skip CUDA attention regressions when no device is visible

2026-10-02 22:54:32 +08:00

core

[Memory] Reduce SM70 graph scratch and DFlash active peak

2026-09-23 09:58:34 +08:00

helion

…

ir

…

mamba

[Bugfix] Materialize accepted Mamba state before single-token tails

2026-09-21 20:04:31 +08:00

moe

[Kernel] Select exact router top keys and retain HC schedule controls

2026-09-27 16:44:31 +08:00

quantization

[Kernel] Improve SM70 M48 supply, long softmax and TP4 reduction

2026-09-27 18:16:21 +08:00

__init__.py

…

allclose_default.py

…

quant_utils.py

…

test_apply_repetition_penalties.py

…

test_awq_int4_to_int8.py

…

test_cache_kernels.py

…

test_compressor_kv_cache.py

…

test_concat_mla_q.py

…

test_cp_gather_fp8.py

…

test_deepseek_v4_sm70_fp8_software.py

…

test_deepseek_v4_sm70_indexer.py

…

test_deepseek_v4_sm70_qnorm_rope_kv_insert.py

…

test_fla_layernorm_guard.py

…

test_flex_attention.py

…

test_fused_deepseek_v4_qnorm_rope_kv_insert.py

…

test_fused_gdn_post_conv.py

…

test_fused_indexer_q_rope_quant.py

…

test_fused_inv_rope_fp8_quant.py

…

test_fused_quant_activation.py

…

test_fused_recurrent_packed_decode.py

…

test_fused_sigmoid_gating_delta_rule.py

[Kernel] Preserve q8 GDN numerics with strided packed verification

2026-09-08 09:12:11 +08:00

test_kda.py

…

test_kpool_decode_update_batched.py

…

test_mhc_fp16_range.py

[Bugfix][DeepSeek-V4] Keep mHC finite under float16 for attention-sink rows

2026-09-19 16:39:49 +02:00

test_mhc_kernels.py

…

test_mhc_sm70_fp16.py

…

test_onednn.py

…

test_shuffle_rows.py

…

test_sm70_awq_chunked_w2.py

[Test][SM70] Cover compact W2 metadata and distributed route graph replay

2026-09-07 12:50:40 +08:00

test_sm70_awq_qpn_m1.py

…

test_sm70_deepseek_v4_fp16_gemv.py

…

test_sm70_dflash2_packed_gdn_fp32.py

[Kernel] Extend exact SM70 q8 GDN tiles to intermediate batches

2026-09-27 13:14:06 +08:00

test_sm70_hc_batch_reuse_contract.py

[Bench] Screen exact native HC batch projections and TP4 chain

2026-09-27 09:10:34 +08:00

test_sm70_mtp_gdn_mixed_qkv.py

[Kernel] Reuse exact router and GDN loaders for MTP batches

2026-09-27 09:17:50 +08:00

test_sm70_mtp_ple_conv.py

[Kernel] Fuse exact MTP4 PLE rollback and convolution

2026-09-27 18:03:35 +08:00

test_sm70_mtp_router_batch.py

[Kernel] Fuse MTP batch HC launches and router split reduction

2026-09-27 14:57:05 +08:00

test_sm70_mtp_shared_batch.py

[Kernel] Fuse exact MTP shared expert batch epilogues

2026-09-27 17:35:44 +08:00

test_sm70_qpn_concurrency.py

[Kernel][SM70] Integrate concurrent QPN projection and verifier improvements

2026-09-07 18:38:08 +08:00

test_sm70_qsa_page4_plan.py

…

test_sm70_qwen35_gdn_split.py

[Kernel] Materialize combined DFlash2 GDN tails in one launch

2026-09-08 12:20:14 +08:00

test_sm70_rmsnorm_gated_exact.py

[Bugfix] Pin Qwen3.8 gated norm to native FP32 arithmetic

2026-09-28 00:48:57 +08:00

test_sm70_shared_gate_epilogue.py

[Kernel][WIP] Extend Qwen3.8 HC and FP16 input fast paths to batch decode

2026-09-27 10:50:33 +08:00

test_top_k_per_row.py

[Kernel] Unroll cooperative HC and reuse exact QSA batch selection

2026-09-27 16:25:15 +08:00

utils.py

…