yangzhuxinyzx fc3901beb4 [Core] Share SM70 AWQ and FP8 MoE stages and capture engine policy ...

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

2026-10-09 10:54:32 +08:00

..

cpu

…

deepgemm

…

ir

…

__init__.py

…

bench_concat_mla_q.py

…

bench_cp_gather_fp8.py

…

benchmark_2d_silu_mul_fp8_quant.py

…

benchmark_activation.py

…

benchmark_block_fp8_gemm.py

…

benchmark_cutlass_moe_fp8.py

…

benchmark_cutlass_moe_nvfp4.py

…

benchmark_deepseek_v4_sm70_sparse_decode.py

…

benchmark_device_communicators.py

…

benchmark_dflash2_lookup.py

…

benchmark_dspark_markov.py

…

benchmark_dspark_metadata.py

…

benchmark_dspark_sm70_attention.py

…

benchmark_flash_v100_bfla_prefill.py

…

benchmark_flash_v100_dflash2_swa.py

…

benchmark_flash_v100_fp8_kv.py

…

benchmark_flash_v100_prefill_decay.py

…

benchmark_flashinfer_decode_sm120.py

…

benchmark_flashinfer_prefill_sm120.py

…

benchmark_fp8_gemm.py

…

benchmark_fused_collective.py

…

benchmark_fused_moe_lora_one_shot.py

…

benchmark_fused_topk.py

…

benchmark_gguf_dense_segment_banks.py

[Kernel] Cover replicated KV head projection segments

2026-10-06 00:42:10 +08:00

benchmark_gguf_dp4a_down.py

[Kernel] Select calibrated integer GGUF expert graphs by capability

2026-10-05 17:44:43 +08:00

benchmark_gguf_dp4a_expert.py

[Kernel] Fuse GGUF integer down projection and route reduction

2026-10-05 17:32:25 +08:00

benchmark_gguf_fp16_ba.py

[Kernel] Keep GDN a/b row arithmetic consistent through M32

2026-10-04 20:44:51 +08:00

benchmark_gguf_fp16_projection.py

[Test] Validate floating GGUF shards in normal graph runtime

2026-10-05 13:30:21 +08:00

benchmark_gguf_iq3_gated.py

[Doc] Record packaged GGUF checks and main baseline

2026-10-05 12:45:19 +08:00

benchmark_gguf_iq4_native.py

[Kernel] Preserve IQ4_XS source bytes and nested scales (#950)

2026-10-05 11:11:01 +08:00

benchmark_gguf_lut4_experts.py

[Doc] Record matched IQ4 model quality and partition screening

2026-10-06 05:47:14 +08:00

benchmark_gguf_native_linear.py

[Kernel] Read prepared affine streams on the shared-activation skeleton

2026-10-05 23:30:08 +08:00

benchmark_gguf_native_pair.py

[Kernel] Admit measured pure IQ3_XXS pairs with merged canonical fallback

2026-10-06 01:03:43 +08:00

benchmark_gguf_ordered_rows.py

[Doc] Validate ordered PLE row decoding and graph cost

2026-10-05 15:39:53 +08:00

benchmark_gguf_pair_reader_cursor.py

[Kernel] Share activations across IQ3_S and IQ4_XS gated pairs

2026-10-05 13:15:31 +08:00

benchmark_gguf_pinned_ple_numa.py

[Kernel] Read packed GGUF PLE decode rows from pinned TP shards (#996)

2026-10-06 04:00:12 +08:00

benchmark_gguf_q8_intermediate.py

[Kernel] Retain sixteen-lane routed integer schedule

2026-10-05 22:45:16 +08:00

benchmark_gguf_qkv.py

[Kernel] Wire GGUF small-M attention inputs by measured capability

2026-10-06 04:09:09 +08:00

benchmark_gguf_qkvz.py

[Kernel] Wire measured joint GDN inputs with canonical stream aliases

2026-10-06 01:44:44 +08:00

benchmark_gguf_raw_grouped.py

[Doc] Qualify standalone joint experts with cold-bank timings

2026-10-05 06:10:46 +08:00

benchmark_gguf_sm70.py

[Kernel] Package native GGUF formats and Volta prefill routing

2026-10-03 09:45:41 +08:00

benchmark_gguf_small_down.py

[Bench] Measure grouped down vectors with CUDA graph replay

2026-10-05 01:08:54 +08:00

benchmark_gguf_small_outputs.py

[Kernel] Admit measured shared-activation output projections

2026-10-06 03:50:09 +08:00

benchmark_gguf_turbomind_grouped.py

[Bench] Identify loaded GGUF core and retain rejected IQ3 evidence

2026-10-04 03:27:23 +08:00

benchmark_gguf_turbomind.py

[Bench] Identify loaded GGUF core and retain rejected IQ3 evidence

2026-10-04 03:27:23 +08:00

benchmark_glm53_fp16_gemv.py

[Kernel] Speed up exact GLM verifier GEMV

2026-09-03 06:39:06 +08:00

benchmark_grouped_gemm_cutlass.py

…

benchmark_h3_vsa_fp32.py

[Benchmark] Preserve the H3 VSA FP32 quality diagnostic and stage results

2026-09-10 05:15:43 +08:00

benchmark_int8_gemm.py

…

benchmark_layernorm.py

…

benchmark_lora.py

…

benchmark_machete.py

…

benchmark_marlin.py

…

benchmark_mla_k_concat.py

…

benchmark_moe_align_block_size.py

…

benchmark_moe_defaults.py

…

benchmark_moe_permute_unpermute.py

…

benchmark_moe.py

…

benchmark_mrope.py

…

benchmark_mxfp4_qutlass.py

…

benchmark_nvfp4_gemm.py

…

benchmark_nvfp4_quant.py

…

benchmark_nvfp4_qutlass.py

…

benchmark_paged_attention.py

…

benchmark_per_token_group_quant.py

…

benchmark_per_token_quant_fp8.py

…

benchmark_quant.py

…

benchmark_qwen38_fp16_hc_mix_mtp5.py

[Kernel][SM70] Add Qwen3.8 MTP5 verifier fast paths

2026-08-28 19:37:31 +08:00

benchmark_qwen38_fp16_verifier_m5.py

[Kernel][WIP] Add SM70 batch MoE candidates and raw-scale safety gates

2026-09-05 11:36:33 +08:00

benchmark_qwen38_nvfp4_qpn_mtp5.py

[Bugfix][SM70] Gate MTP5 NVFP4 by explicit capability

2026-09-03 19:43:20 +08:00

benchmark_qwen38_qsa_mtp5_flash_v100.py

[Kernel][WIP] Add SM70 batch MoE candidates and raw-scale safety gates

2026-09-05 11:36:33 +08:00

benchmark_qwen38_router_mtp5.py

[Kernel][WIP] Add SM70 batch MoE candidates and raw-scale safety gates

2026-09-05 11:36:33 +08:00

benchmark_reshape_and_cache_flash.py

…

benchmark_reshape_and_cache.py

…

benchmark_rmsnorm.py

…

benchmark_rope.py

…

benchmark_selective_state_update.py

…

benchmark_shapes.py

…

benchmark_silu_mul_fp8_quant.py

…

benchmark_sm70_compact_half_topk.py

[Kernel] Reduce SM70 target compact top-k launches (#942)

2026-10-05 06:55:03 +08:00

benchmark_sm70_compact_top1_ipc.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_dflash2_grouped_verify.py

…

benchmark_sm70_dflash2_qk_rope.py

[Kernel] Fuse SM70 DFlash2 query preparation and attention layouts (#935)

2026-10-05 06:25:52 +08:00

benchmark_sm70_dflash_batched_prefill.py

[Core] Enable measured SM70 DFlash batch defaults

2026-09-25 21:33:26 +08:00

benchmark_sm70_draft_f16_layout.py

[Perf] Audit strict draft GEMM and stage TP4 GDN value tiling

2026-09-09 12:24:28 +08:00

benchmark_sm70_exact_row_reduce.py

[Kernel] Add explicit calibrated SM70 local-row reduction

2026-09-09 21:33:26 +08:00

benchmark_sm70_flash_v100_xqa_layout.py

…

benchmark_sm70_fp8_e5m2_cache_write.py

…

benchmark_sm70_fp8_graph_capture.py

…

benchmark_sm70_fp16_gemv_silu.py

[Kernel] Improve small-grid FP16 GEMV SiLU scheduling on SM70 (#812)

2026-10-03 11:06:40 +08:00

benchmark_sm70_gdn_input_batch.py

[Merge] Integrate MTP kernels without changing concurrent arithmetic

2026-09-28 00:43:59 +08:00

benchmark_sm70_gdn_preprocess.py

[Bench] Compare a 36-layer GDN preprocessing chain under CUDA graphs

2026-10-05 02:34:02 +08:00

benchmark_sm70_gdn_projection_split.py

[Kernel][SM70] Fuse copy-only batched GDN projection splits

2026-09-07 18:14:47 +08:00

benchmark_sm70_gdn_projection_tails.py

[Bench] Record GDN tail-copy graph timing and bitwise validation

2026-10-05 03:14:57 +08:00

benchmark_sm70_gguf_head.py

[Kernel] Extend the measured canonical LM-head route to Q6_K

2026-10-04 20:13:05 +08:00

benchmark_sm70_grouped_attention_long.py

[Bench] Validate visible q8 rounds and isolate tail attention costs

2026-09-10 09:44:09 +08:00

benchmark_sm70_grouped_attention_precision.py

[Bench] Specialize q8 visibility and audit native pre-cast precision

2026-09-10 07:57:09 +08:00

benchmark_sm70_grouped_attention_qk_stage.py

[Bench] Screen exact QK reuse toward the 7-ms attention target

2026-09-10 14:12:05 +08:00

benchmark_sm70_grouped_attention_splits.py

[Perf] Measure long-context increments and reject slower schedules

2026-09-10 01:17:39 +08:00

benchmark_sm70_hc_batch_native_tp4.py

[Merge] Integrate MTP kernels without changing concurrent arithmetic

2026-09-28 00:43:59 +08:00

benchmark_sm70_hc_batch_reuse.py

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

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

benchmark_sm70_hc_batch_tp4.py

[Perf][WIP] Gate HC sharding on rotating-weight performance

2026-09-05 15:08:29 +08:00

benchmark_sm70_hc_batch.py

[Perf][WIP] Screen batched FP16 HC fusion and TP4 sharding

2026-09-05 14:19:59 +08:00

benchmark_sm70_hc_cta_chain.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_hc_full_chain.py

[Kernel] Fuse exact SM70 HC up projection and output gather

2026-09-05 15:21:13 +08:00

benchmark_sm70_hc_gguf_tp4.py

[Bench] Use an unambiguous batch selector under torchrun

2026-10-05 02:10:56 +08:00

benchmark_sm70_hc_ll.py

[Bench] Compare packaged HC shards and replicated kernels in ABBA

2026-10-06 01:37:05 +08:00

benchmark_sm70_hc_tp4.py

[Kernel] Fuse exact SM70 HC up projection and output gather

2026-09-05 15:21:13 +08:00

benchmark_sm70_lm_head_local_vocab.py

[Kernel][WIP] Generalize SM70 attention and projection shard layouts

2026-09-21 11:41:57 +08:00

benchmark_sm70_long_attention_batch.py

[Kernel] Extend SM70 long-context q8 attention to batches

2026-09-27 01:46:18 +08:00

benchmark_sm70_long_context_attention.py

[Bugfix] Reject invalid attention benchmark library loaders

2026-10-05 02:03:36 +08:00

benchmark_sm70_moe_packed_w13.py

[Kernel][WIP] Integrate opt-in grouped SM70 MoE decode

2026-09-05 13:35:53 +08:00

benchmark_sm70_mtp5_grouped.py

[Kernel] Qualify exact Flash-Next MTP4 batch decode stack

2026-09-27 02:41:08 +08:00

benchmark_sm70_mtp_current_chains.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_mtp_fc_gather_pair.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_mtp_fp32_dense.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_mtp_gdn_output.py

[Bench] Record direct MTP recurrent output copy savings

2026-10-05 03:40:28 +08:00

benchmark_sm70_mtp_local_top1.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_mtp_moe_fp16.py

[Kernel] Add exact SM70 FP16 draft MoE projections

2026-09-27 03:24:00 +08:00

benchmark_sm70_mtp_qpn8_dense.py

[Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821)

2026-10-05 19:59:24 +08:00

benchmark_sm70_mtp_strided_qkv.py

[Bench] Record strided MTP QKV timing and exact state validation

2026-10-05 03:25:52 +08:00

benchmark_sm70_multihead_attention.py

[Kernel][WIP] Generalize SM70 attention and projection shard layouts

2026-09-21 11:41:57 +08:00

benchmark_sm70_ngram_input.py

[Bench] Validate PLE preparation with actual pinned token storage

2026-10-05 22:51:56 +08:00

benchmark_sm70_nvfp4_bundle.py

[Kernel] Bundle native SM70 NVFP4 codes and scales

2026-10-06 22:37:40 +08:00

benchmark_sm70_nvfp4_qpn2.py

Revert "[Core][SM70] Trace DFlash2 concurrency boundaries"

2026-09-04 18:36:23 +08:00

benchmark_sm70_nvfp4_shared_weight.py

[Kernel][WIP] Generalize SM70 attention and projection shard layouts

2026-09-21 11:41:57 +08:00

benchmark_sm70_packed_topk_gather.py

[Kernel] Gather compact SM70 vocabulary candidates in one message

2026-10-08 05:43:32 +08:00

benchmark_sm70_ple_ngram_batch.py

[Bench] Compare exact PLE n-gram routes under CUDA graph replay

2026-10-04 23:26:09 +08:00

benchmark_sm70_qpn2_chunked.py

[Perf] Validate bound attention and screen chunked QPN2

2026-09-09 01:30:59 +08:00

benchmark_sm70_qpn2_cooperative_mlp.py

[Perf] Add cooperative MLP screen and isolate GDN round savings

2026-09-09 14:31:15 +08:00

benchmark_sm70_qpn2_graph_layout.py

[Perf] Add QPN2 layout checks on unchanged CUDA graphs

2026-09-09 16:13:49 +08:00

benchmark_sm70_qpn2_packed_mlp.py

[Perf] Add reproducible QPN2 packed-boundary experiments

2026-09-08 16:54:59 +08:00

benchmark_sm70_qpn2_publish.py

[Perf] Make QPN2 publication experiments reproducible

2026-09-08 14:11:11 +08:00

benchmark_sm70_qpn2_working_set.py

[Perf] Record exact q8 QPN2 scheduling gains

2026-09-08 18:16:46 +08:00

benchmark_sm70_qsa_batch_decode.py

[Kernel][WIP] Screen exact C2 selector and fused dense reductions

2026-09-27 13:07:23 +08:00

benchmark_sm70_quasar_nvfp4_oracle.py

[Bugfix] Guard DFlash2 sampling cutoffs and repair operator audit oracles

2026-09-06 02:46:56 +08:00

benchmark_sm70_quasar_tp_quality.py

[Audit] Trace QUASAR TP partition drift through target sampling

2026-09-06 14:01:26 +08:00

benchmark_sm70_qwen38_batch_gate.py

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

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

benchmark_sm70_qwen38_dense_batch.py

[Kernel][WIP] Screen exact C2 selector and fused dense reductions

2026-09-27 13:07:23 +08:00

benchmark_sm70_ring_atomic_allreduce.py

[Bench] Retry collective screens while foreign GPU work remains

2026-10-04 14:00:43 +08:00

benchmark_sm70_rmsnorm_gated_exact.py

[Kernel] Qualify exact Flash-Next MTP4 batch decode stack

2026-09-27 02:41:08 +08:00

benchmark_sm70_scalar_attention.py

[Bench] Expose exact long q1 scalar reuse and record paired gains

2026-09-10 06:08:23 +08:00

benchmark_sm70_small_moe_routing.py

[Bench] Keep routing graph timing above replay launch gaps

2026-10-04 21:49:40 +08:00

benchmark_sm70_sparse_dense_topk.py

[Perf] Add reproducible sparse dense-order top-k screen

2026-09-08 23:43:38 +08:00

benchmark_sm70_tp2_e4m3_scalar.py

[Merge] Audit TP2 endpoint against current main

2026-09-10 00:19:29 +08:00

benchmark_sm70_tp4_norm_packets.py

[Kernel] Reuse M8 gate operands and TP4 norm packets (#1041)

2026-10-08 08:50:20 +08:00

benchmark_sm70_tp4_small_message_push.py

[Perf][WIP] Gate HC sharding on rotating-weight performance

2026-09-05 15:08:29 +08:00

benchmark_trtllm_decode_attention.py

…

benchmark_trtllm_prefill_attention.py

…

benchmark_vit_bilinear_pos_embed.py

…

benchmark_vit_fp8_attn.py

…

benchmark_w8a8_block_fp8.py

…

build_sm70_exact_decode_screen.py

[Kernel] Reuse M8 gate operands and TP4 norm packets (#1041)

2026-10-08 08:50:20 +08:00

build_sm70_grouped_attention_candidate.py

[Bench] Screen exact DFlash2 attention resource changes

2026-09-10 11:48:51 +08:00

build_sm70_grouped_attention_compact_cta.py

[Bench] Screen exact QK reuse toward the 7-ms attention target

2026-09-10 14:12:05 +08:00

build_sm70_grouped_attention_direct_qk.py

[Bench] Screen exact QK reuse toward the 7-ms attention target

2026-09-10 14:12:05 +08:00

build_sm70_grouped_attention_n64.py

[Bench] Validate visible q8 rounds and isolate tail attention costs

2026-09-10 09:44:09 +08:00

build_sm70_grouped_attention_phase_probe.py

[Kernel] Screen compact probability layout for DFlash2 attention

2026-09-10 12:38:02 +08:00

build_sm70_grouped_attention_pipeline.py

[Bench] Expose exact long q1 scalar reuse and record paired gains

2026-09-10 06:08:23 +08:00

build_sm70_grouped_attention_qk_parallel.py

[Bench] Screen exact DFlash2 attention resource changes

2026-09-10 11:48:51 +08:00

build_sm70_grouped_attention_staged.py

[Bench] Validate scalar E4M3 lookup and attribute ordered attention phases

2026-09-10 07:36:20 +08:00

build_sm70_grouped_attention_swizzle.py

[Bench] Screen exact QK reuse toward the 7-ms attention target

2026-09-10 14:12:05 +08:00

build_sm70_long_context_attention.py

[Benchmark] Check E4M3 decoder equivalence and archive revisions

2026-10-05 02:03:36 +08:00

build_sm70_native_sort_candidate.py

[Perf] Add reproducible sparse dense-order top-k screen

2026-09-08 23:43:38 +08:00

build_sm70_qpn2_chunked_candidate.py

[Perf] Validate bound attention and screen chunked QPN2

2026-09-09 01:30:59 +08:00

build_sm70_qpn2_cooperative_mlp.py

[Perf] Add cooperative MLP screen and isolate GDN round savings

2026-09-09 14:31:15 +08:00

build_sm70_qpn2_dual_norm.py

[Perf] Add QPN2 layout checks on unchanged CUDA graphs

2026-09-09 16:13:49 +08:00

build_sm70_qpn2_packed_input_candidate.py

[Perf] Add reproducible QPN2 packed-boundary experiments

2026-09-08 16:54:59 +08:00

build_sm70_qpn2_publish_candidate.py

[Perf] Add reproducible QPN2 packed-boundary experiments

2026-09-08 16:54:59 +08:00

build_sm70_qpn2_q8_candidate.py

[Perf] Record exact q8 QPN2 scheduling gains

2026-09-08 18:16:46 +08:00

build_sm70_qsa_topk_sidecar.py

[Kernel] Optimize lossless M1 router keys and QSA addressing

2026-09-05 19:30:09 +08:00

build_sm70_qwen38_runtime.py

[Bench] Reproduce the Qwen3.8 SM70 single-request baseline from source

2026-09-06 19:43:16 +08:00

build_sm70_scalar_attention_candidate.py

[Bench] Validate visible q8 rounds and isolate tail attention costs

2026-09-10 09:44:09 +08:00

build_sm70_tp2_matched_qpn2.py

[Kernel] Add a reproducible TP2 matched QPN2 build

2026-09-09 15:25:28 +08:00

build_sm70_w13_exact_sidecar.py

[Kernel][WIP] Screen exact W13 load and partial scheduling

2026-09-06 00:51:11 +08:00

deepseek_v4_sm70_aot_check.py

…

dflash2_split_attention_reference.py

[Kernel] Parallelize DFlash2 small-query window attention on SM70 (#944)

2026-10-05 08:13:09 +08:00

graph_machete_bench.py

…

h3_vsa_fp32.cu

[Benchmark] Batch independent H3 FP32 sparse queries

2026-09-10 05:35:02 +08:00

requirements.txt

…

sm70_context_probe_candidate_route.py

[Perf] Stage draft context behind the target cutoff probe

2026-09-09 01:58:27 +08:00

sm70_dflash2_common_candidate_route.py

[Perf] Enforce the frozen sparse-sort runtime contract

2026-09-09 19:31:54 +08:00

sm70_dflash2_qpn2_candidate_route.py

[Perf] Package optional QPN2 routes beside common DFlash2 schedules

2026-09-09 18:39:52 +08:00

sm70_draft_column_candidate_route.py

[Perf] Audit strict draft GEMM and stage TP4 GDN value tiling

2026-09-09 12:24:28 +08:00

sm70_gdn_value_tile_candidate_route.py

[Perf] Audit strict draft GEMM and stage TP4 GDN value tiling

2026-09-09 12:24:28 +08:00

sm70_grouped_attention_candidate_route.py

[Perf] Audit full-round resources and bind actual attention module

2026-09-09 00:24:18 +08:00

sm70_hc_batch_screen.cu

[Perf][WIP] Screen batched FP16 HC fusion and TP4 sharding

2026-09-05 14:19:59 +08:00

sm70_moe_stage_parity.py

[Core] Share SM70 AWQ and FP8 MoE stages and capture engine policy

2026-10-09 10:54:32 +08:00

sm70_nvfp4_shared_sidecar.cpp

[Kernel] Reduce DFlash2 scale and LM-head residency

2026-09-12 23:45:29 +08:00

sm70_paired_stats.py

[Build] Support SM72 targets and robust device diagnostics

2026-10-03 06:27:04 +08:00

sm70_qpn2_graph_layout.py

[Perf] Add QPN2 layout checks on unchanged CUDA graphs

2026-09-09 16:13:49 +08:00

sm70_qpn2_graph_nodes.py

[Perf] Add QPN2 layout checks on unchanged CUDA graphs

2026-09-09 16:13:49 +08:00

sm70_qsa_topk_sidecar.cu

[CI] Clear NVFP4 integration checks without changing runtime arithmetic

2026-09-06 16:57:36 +08:00

sm70_qwen38_custom_ar_sidecar.cu

[Bench] Reproduce the Qwen3.8 SM70 single-request baseline from source

2026-09-06 19:43:16 +08:00

sm70_qwen38_nvfp4_sidecar.cpp

[Bench] Reproduce the Qwen3.8 SM70 single-request baseline from source

2026-09-06 19:43:16 +08:00

sm70_qwen38_qk_rope_micro.py

[Kernel] Fuse Qwen query preparation and KV publication on SM70

2026-10-04 23:19:15 +08:00

sm70_w13_exact_sidecar.cu

[Bugfix][SM70] Integrate role-specific projection dispatch with main

2026-09-07 11:10:12 +08:00

utils.py

…

verify_sm70_hc_down_scatter.py

[CI] Clear NVFP4 integration checks without changing runtime arithmetic

2026-09-06 16:57:36 +08:00

verify_sm70_nvfp4_moe_raw_storage.py

[Kernel][WIP] Integrate opt-in grouped SM70 MoE decode

2026-09-05 13:35:53 +08:00

verify_sm70_output_projection_policy.py

[Bugfix][SM70] Integrate role-specific projection dispatch with main

2026-09-07 11:10:12 +08:00

verify_sm70_qsa_nvfp4_relocation.py

[Test] Integrate QSA allocation repair with NVFP4 causal regression

2026-09-06 15:04:03 +08:00

verify_sm70_qsa_resolved.py

[CI] Clear NVFP4 integration checks without changing runtime arithmetic

2026-09-06 16:57:36 +08:00

verify_sm70_qsa_router_exact.py

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

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

verify_sm70_w13_exact.py

[Bugfix][SM70] Integrate role-specific projection dispatch with main

2026-09-07 11:10:12 +08:00

weight_shapes.py

…