yangzhuxinyzx a42e887cf6 [Kernel] Improve SM70 M48 supply, long softmax and TP4 reduction ...

Add shared compressed-weight M48 and tail tiles with partial-warp guards. Preserve reference split-K selection, attention reduction trees and rank sum order. Record exact-token ordinary service gains and the remaining C4/C6/C8 performance targets.

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

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

..

nvfp4_utils.py

…

test_allspark_gemm.py

…

test_awq_triton.py

…

test_awq.py

…

test_block_fp8.py

…

test_block_int8.py

…

test_cpu_fp8_scaled_mm.py

…

test_cutlass_scaled_mm.py

…

test_cutlass_w4a8_moe.py

…

test_cutlass_w4a8.py

…

test_flashinfer_nvfp4_scaled_mm.py

…

test_flashinfer_scaled_mm.py

…

test_fp8_min_max_helper.py

…

test_fp8_quant_group.py

…

test_fp8_quant.py

…

test_ggml.py

…

test_gguf.py

…

test_gptq.py

…

test_hadacore.py

…

test_int8_kernel.py

…

test_int8_quant.py

…

test_machete_mm.py

…

test_marlin_gemm.py

…

test_mxfp4_qutlass.py

…

test_mxfp4_triton_ep.py

…

test_nvfp4_emulation.py

…

test_nvfp4_quant.py

…

test_nvfp4_qutlass.py

…

test_nvfp4_scaled_mm.py

…

test_per_token_group_quant.py

…

test_rdna3_w4a16_selection.py

…

test_rdna3_w4a16.py

…

test_rocm_compressed_tensors_w4a16.py

…

test_rocm_skinny_gemms.py

…

test_scaled_mm_kernel_selection.py

…

test_silu_mul_nvfp4_quant.py

…

test_sm70_dflash_context_reduction.py

[Kernel] Speed up SM70 concurrent long decode

2026-09-27 07:48:20 +08:00

test_sm70_marlin_splitk.py

…

test_sm70_moe_packed_w13.py

…

test_sm70_mxfp4_e8m0.py

…

test_sm70_nvfp4_output_padding.py

…

test_sm70_nvfp4_prescaled_batch.py

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

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

test_sm70_nvfp4_qpn2_scales.py

[Bugfix] Preserve shared score dependencies across CUDA graphs

2026-09-22 10:02:19 +08:00

test_sm70_nvfp4_raw_scales.py

…

test_sm70_prefill_local_layouts.py

…

test_sm70_qpn2_m32_packed_input.py

[Kernel] Enable shared SM70 C2 C4 C8 batch acceleration by default (#691)

2026-09-26 18:43:15 +08:00

test_sm70_qpn8_local_layouts.py

…

test_sm70_qpn8_m32_packed_input.py

[Kernel] Enable shared SM70 C2 C4 C8 batch acceleration by default (#691)

2026-09-26 18:43:15 +08:00

test_sm70_quantized_batch_supply.py

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

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

test_triton_scaled_mm.py

…

test_triton_w4a16.py

…

test_w4a16_kernel_selection.py

…