fbtriton v3.7.3

fbtriton v3.7.3

AMD-perf point release on release/3.7.x, targeting gfx950 / MI350. v3.7.2 was fixes-only and dropped the AMD perf work; v3.7.3 brings it onto the production release line.

Highlights (AMD / MI350, gfx950)

• Paged-attention decode — page-16 64-token tile aggregation, pipelined page-64, wide async tiles + tuned split-K.
• Dense a16w16 GEMM — 8-wave gfx950 GEMM with pinned layouts, auto split-K for skinny shapes, thin-N tiling, and a coalesced dwordx4 FP16 epilogue.
• MXFP4 / w4a4 MoE GEMM — new intra_wave and inter_wave 8-wave a4w4 kernels (packed E2M1 / E8M0 scales, FP32 accumulation, BF16 output).
• addmm + GLU epilogue fusion.
• Supporting TLX layout / warp-spec infrastructure (user-pinned layouts, 256-bit loads, layout propagation/auto-release).

Validation (gfx950 / MI350)

• TLX correctness: 69 passed / 4 skip; a4w4 intra + inter correctness green; folded-chain LIT green.
• Perf (clock-locked at 2100 MHz) vs from-source main: parity across target cases — paged decode geomean 0.995×, a16w16 0.96–1.01×, MXFP4 a4w4 K≥2048 1.00–1.02×. No regression vs v3.7.2 on shared paths.

Notes

• Version: 3.7.3+fb.
• Deferred: torchTLX Inductor (Chain C) and the modulo scheduler (Chain D).
• Full pick/shim log, dependency DAG, and perf tables: .backports/3.7.3/backport_3.7.3.md.
• Backport PR: #2335.

🤖 Backport prepared with Claude Code.