yangzhuxinyzx 7413fad39a [Kernel] Enable qualified full-vocabulary Flash-Next MTP4 defaults (#821) ...

Enable the qualified full-vocabulary Flash-Next MTP4 defaults: draft-only QPN8 head, eight-warp HC, FP32 shared-expert partials, compact local top1/IPC, one FC gather and disk PLE. Keep shared numerical admission tools and bounded sampler scratch. Exclude unadmitted structural and vocabulary experiments.

Validation: normal SM70 source build, 205 CPU tests, 82 focused GPU cases, 30 local top1 oracles and real TP0 head packets. Recorded complete round is 21.198709 ms; the 15-ms target and C4 qualification remain open.

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

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

..

__init__.py

…

conftest.py

…

eplb_utils.py

…

test_ca_buffer_sharing.py

…

test_comm_ops.py

[Core] Cleanup KVConnector handling with PP + fix MRV2 (#43732)

2026-05-28 13:12:03 -07:00

test_context_parallel.py

…

test_custom_all_reduce_allocator.py

[Bugfix][Distributed] Make custom allreduce graph capture VMM-safe (#262)

2026-08-24 23:54:06 +08:00

test_custom_all_reduce_dispatch.py

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

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

test_custom_all_reduce.py

[Core][SM70] Repair Qwen4Exp PLE and AMD integration

2026-08-28 11:53:21 +08:00

test_dcp_a2a.py

[Perf][QSA] Cut DCP decode overhead of the sparse attention layers

2026-10-03 13:10:45 +08:00

test_distributed_oot.py

…

test_elastic_ep.py

…

test_eplb_algo.py

…

test_eplb_events.py

…

test_eplb_execute.py

…

test_eplb_fused_moe_layer_dep_nvfp4.py

[MoE Refactor] EPLB refactoring for FusedMoE (#41055)

2026-05-12 14:16:31 -04:00

test_eplb_fused_moe_layer.py

…

test_eplb_spec_decode.py

…

test_eplb_utils.py

…

test_events.py

…

test_expert_parallel.py

…

test_expert_placement.py

[MoE Refactor] Move expert map related code into ExpertMapManager class (#41046)

2026-05-12 09:18:27 -04:00

test_group_coordinator_timeout.py

[Bugfix] Preserve NCCL timeout for PP sibling groups

2026-09-26 20:32:14 +08:00

test_kv_cache_events.py

feat(kv-events): emit KV cache metadata (#40984)

2026-05-12 15:58:48 +00:00

test_kvlayout.py

…

test_mnnvl_alltoall.py

[Bugfix][MoE] FlashInfer one-sided: workspace union across heterogeneous layers (#42976)

2026-05-19 14:43:04 -04:00

test_mq_connect_ip.py

…

test_multi_node_assignment.py

…

test_multiproc_executor.py

…

test_nccl_graph_register_default.py

[Bugfix][Distributed] Disable NCCL graph buffer registration under expandable segments

2026-10-02 21:07:44 +08:00

test_nccl_symm_mem_allreduce.py

…

test_node_count.py

…

test_packed_tensor.py

[Feat][RL] IPC weight sync optimizations: multigpu support and chunked packed tensors (#37476)

2026-05-15 22:53:06 +08:00

test_pipeline_parallel.py

Deprecate JAISLMHeadModel (#43784)

2026-05-28 18:29:12 +00:00

test_pipeline_partition.py

[Core][Qwen4Exp] Gate PLE under PP on the partition, not the PP size

2026-09-05 20:11:51 +02:00

test_pp_cudagraph.py

…

test_pynccl.py

…

test_quick_all_reduce.py

[ROCm][CI] Extend ROCm quick reduce coverage (#40990)

2026-05-26 21:57:13 +08:00

test_ray_v2_executor_e2e.py

…

test_ray_v2_executor.py

…

test_rocm_quick_reduce.py

[ROCm][CI] Extend ROCm quick reduce coverage (#40990)

2026-05-26 21:57:13 +08:00

test_same_node.py

…

test_shm_broadcast.py

…

test_shm_buffer.py

…

test_shm_storage.py

…

test_sm70_medium_all_reduce.py

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

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

test_sm70_ring_capabilities.py

[Kernel] Guard ring initialization and retain rejected collective dispatch

2026-10-04 13:51:38 +08:00

test_sm70_tp4_push_gemma_norm.py

[Test] Exercise repeated fused norm packet generations

2026-10-05 12:42:00 +08:00

test_sm70_tp8_hierarchical_custom_all_reduce.py

[Kernel] Qualify GLM5 DFlash2 TP8 verifier

2026-09-02 19:10:24 +08:00

test_static_tensor_transfer.py

[Core][Kernel][SM70] Default validated PP2 TP4 paths (#283)

2026-08-26 03:05:28 +08:00

test_symm_mem_allreduce.py

…

test_torchrun_example_moe.py

[Examples] Resettle features examples. (#40995)

2026-04-28 00:33:41 -07:00

test_torchrun_example.py

[Examples] Resettle features examples. (#40995)

2026-04-28 00:33:41 -07:00

test_utils.py

…

test_weight_transfer.py

Add model to WeightTransferEngine.__init__ (#42922)

2026-05-22 17:52:15 -07:00