| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
docs: add AGENTS.md reference to README and simplify agent table | 3 个月前 | |
chore: make agent tooling assistant-agnostic for any AI coding assistant Make the in-repo agent tooling work with any AI coding assistant instead of being tied to a specific vendor, so contributors can use whichever assistant they prefer. | 4 个月前 | |
chore: make agent tooling assistant-agnostic for any AI coding assistant Make the in-repo agent tooling work with any AI coding assistant instead of being tied to a specific vendor, so contributors can use whichever assistant they prefer. | 4 个月前 | |
test(shard-ops): declarative OpShardCase framework + full ST migration Replace the old per-op, per-mesh hand-written shard ST tests with a single declarative framework and migrate every case to it across all three backends. Framework (tests/shard_ops/framework/) - Platform-agnostic core + per-backend adapters (torch cpu/npu, mindspore npu) + per-op declarative case files (cases/case_{op}.py). - 7-step execution: build -> derive -> reference(single-card) -> distribute -> run(multi-card) -> gather -> compare. - OpShardCase fields: inputs/placements/kwargs/extra_inputs, derived_inputs (pre-compute stats on full tensors then slice, for attention sm_max/sm_sum), compare (equal/allclose/shape), compare_outputs (tuple-output indices), needs_mesh (inject mesh for MC2 ops), tags (route cpu/npu x level0/level1), solo_launcher (escape hatch for MC2 launcher isolation). - Placement convention: tuple length == mesh ndim (one entry per mesh axis), Shard(d) shards tensor dim d on that axis. Migrated scope - Torch: 248 cases (201 level0 + 47 level1), CPU(gloo)+NPU(hccl); covers elementwise/reduce/view/index/embedding/matmul/softmax/attention (npu_fusion_attention)/MC2(all_gather_matmul, matmul_reduce_scatter)/... - MindSpore: 238 cases (105 level0 + 133 level1), NPU; mint/mint.nn.functional preferred, DFunction DSA/MHC ops (flash_attention_score, mhc_pre_sinkhorn, mhc_post, lightning_indexer grad_kl_loss, sparse_flash_attention, rope, ...). - 1:1 mapping to the old tests recorded in MIGRATION_STATUS; CANN/ACLNN-only failures triaged as D-class. Performance (level0 wall-clock; legacy figures ran only 6 cases via per-case launchers) backend legacy (6 cases) new (full level0) torch cpu 122s 48.7s (201 cases) torch npu 158s 58.8s (201 cases) ms npu 295s 70.8s (105 cases) The new framework runs the ENTIRE level0 suite faster than the legacy one ran just 6 cases: per-launcher startup (~50s, NPU/CANN/HCCL init) is now amortized across a whole batch instead of paid per case. Wall-clock ~= serial batch count x startup. - Canonicalize mesh axis names at bucketing for non-needs_mesh cases (names are cosmetic there) so same-topology cases merge: torch level1 7->3 batches (~192s->~89s). - Drop redundant 8-card 3D MC2 cases (M/N shards are pure concat; K-reduce + collective already covered at 2/4-card), removing the CANN MC2 HcclAllocComResourceByTiling conflict. MS level1 133 cases ~300s. .agent rules/skills/template updated for the new framework (placement convention, fields, run commands, MC2 / axis-name semantics). | 3 个月前 | |
docs(activation): align checkpoint/swap docs with implementation Update API reference, FAQ, and agent skills per review. Exclude activation_checkpoint.md, pipeline_parallel PP swap section, group_swap, and LlamaFactory activation docs from user-facing materials. | 3 个月前 | |
chore: make agent tooling assistant-agnostic for any AI coding assistant Make the in-repo agent tooling work with any AI coding assistant instead of being tied to a specific vendor, so contributors can use whichever assistant they prefer. | 4 个月前 |