Commit d86f337
committed
docs(upstream): file Path C tracker B + C as PRs #2146/#2147
After re-author against current PR #2142 macro shape (the previous probe-
failed drafts targeted a non-existent tileop scheduler hierarchy), both
patches now apply cleanly on jorgecurious metal-gemm-upstream-rebase + #2142
prereq stack.
Filed:
- PR tile-ai/tilelang#2146 (Path C tracker B): fused FP8 scale broadcast
into T.fp8_scaled_matmul K-loop. 16/8 LOC delta in tilelang/language/
fp8_op.py. Closes the 3-6× audiohacking perf gap on FP8 scaled matmul
per the cppmega.mlx Path C consumer at fp8_vecmat_path_c.py.
- PR tile-ai/tilelang#2147 (Path C tracker C): T.BlockScaledLayout.e8m0_k32
+ T.e8m0_to_float DSL primitive. 5 files touched (tilelang/language/
blockscaled_layout.py new, fp8_op.py extended, __init__.py re-export,
metal_quant.py Metal lowering, e8m0 layout test). Unblocks Sparse-MLA
blockscaled Path C QK reducer.
Both stack on PR #2142 (T.fp8_scaled_matmul intrinsic) which stacks on
PR #2130 (jorgecurious base). Independent of each other — different gaps,
different files (B touches the macro body, C adds the layout primitive).
Receipt _filed_prs_2026_05_04.md updated with rows 13-14.
Total filed PRs: 14 (across ml-explore/mlx, apache/tvm, tile-ai/tilelang,
tile-ai/tvm). All OPEN.
Path C tracker A (pipelined_32x32) shipped in commit 3cb6457 + 6746ff9.
Path C tracker B (#2146) and C (#2147) now filed upstream. All three
Path C follow-up entries from docs/upstream/_path_c_blockers_tracker.md
have landing receipts.1 parent edcc523 commit d86f337
7 files changed
Lines changed: 693 additions & 852 deletions
File tree
- docs/upstream
- tilelang_metal_blockscaled_e8m0
- tilelang_metal_fp8_scaled_matmul_fused_scheduler
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
103 | 103 | | |
104 | 104 | | |
105 | 105 | | |
| 106 | + | |
| 107 | + | |
| 108 | + | |
| 109 | + | |
| 110 | + | |
| 111 | + | |
| 112 | + | |
| 113 | + | |
| 114 | + | |
| 115 | + | |
| 116 | + | |
| 117 | + | |
| 118 | + | |
| 119 | + | |
106 | 120 | | |
107 | 121 | | |
108 | 122 | | |
| |||
0 commit comments