| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
xlite: update README.md Signed-off-by: lulina1 <lina.lulina@huawei.com> | 4 个月前 | |
xlite: fix shape/dtype discrepancies between _C.pyi and doc/kernels Cross-checked all 53 doc/kernels/*.md against _C.pyi docstrings, the kernel sources and tests; corrected 33 errors found on either side: _C.pyi (30): - transpose_1_2: drop nonexistent eye parameter (binding has 2 tensors) - linear_att_conv_and_silu: add missing packed-mode params (query_start_loc/query_lens) and document both modes - linear_att_conv_and_silu_token: add missing stub entirely - recurrent_gated_delta_rule: add missing packed-mode optional params - matmul_dequant: output is fp16 (not bf16); deq_scale is flat [2n] - rmsnorm/rmsnorm_variance_only: variance is [tokens,1] fp32 only - dequant: scale is per-token [m] (kernel indexes by row) - msd_merge_dequant: y_merged is interleaved (row 2r/2r+1), not half-stacked; document counts - unpack_activation: output is [2m, n/2] interleaved - beta_decay: clarify A_log is log-space input, not output g - rope_and_cache: n_heads/n_kv_heads are global (host divides by tp) - rope_complex/mla_prepare/indexer_prepare: freqs is always fp32/ complex64, not model dtype - topk: indices is 1-D [max_seq_len]; outIndices rows = sum(query_lens) - mla_v2: topk_indices is [total_query_tokens, top_k] - permutation: counts is full-width [n_routed_experts]; unp_idx final column is segment start offsets; input is bf16 only - unpermutation: input is the permuted buffer [max_expert_sorted, ...] - hc_post: fix transposed comb indexing (comb[m, h*K+k], source-major) - hc_act: eps also added to softmax input - einsum variants: correct transpose-path descriptions (were swapped) - DeepSeek-V4 config: idx_wq_b/idx_weights_proj are TP-sharded; gate_bias optional on sqrtsoftplus layers; se_up_gate_deq_scale sharded form doc/kernels (3): - rope_and_cache: n_heads/n_kv_heads are global, host divides by tpSize - mla_v2: dense mode requires topK>0 (= index_topk), not 0 - recurrent_gated_delta_rule: 7 tensors share dtype, not 8 - transpose_1_2: add missing bfloat16_t variant Co-Authored-By: GLM-5.3 Signed-off-by: lulina1 <lina.lulina@huawei.com> | 3 小时前 | |
xlite: capture and replay AllGather/ReduceScatter (DP) on the main stream | 1 个月前 | |
xlite: automatically generate version from git tags and write to _version.py Co-authored-by: GLM-5.2 Signed-off-by: Sijie Fu <fusijie@huawei.com> | 11 天前 | |
xlite: tests: Update vLLM server port and environment handling in aisbench; add automatic quantization detection Co-authored-by: GLM-5.1 Signed-off-by: Sijie Fu <fusijie@huawei.com> | 2 个月前 | |
xlite: add feature_matrix.md Signed-off-by: lulina1 <lina.lulina@huawei.com> | 1 个月前 | |
xlite: update models.md Signed-off-by: lulina1 <lina.lulina@huawei.com> | 1 个月前 | |
xlite: fix review problems of quant_ops | 5 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 4 个月前 | ||
| 3 小时前 | ||
| 1 个月前 | ||
| 11 天前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 5 个月前 |