[ashhart]

DeepSeek-V4.1 decode on Metal spends about 40% of a layer in the dense FP8 projections (fp8_rows.metal). At one row they read about 675 GB/s on an M5 Ultra, which streams one large read at about 1.1 TB/s. This change gets them closer and keeps every output bit.

• Loads first. The one-launch kernel issues every weight and scale load of a lane before its FMA chains.
• Tile roots in one simdgroup. An output's tile roots now meet in one simdgroup's butterfly. Lane k holds tile k's root, zero past the tiles. That's the same zero-padded adjacent-pair tree one thread used to walk over 32 roots, so the sums and their order don't change. Most of the gain is here.
• Kind d. A new block kind takes one column a simdgroup over 32-group tiles. kindOf picks it for K in whole 1024s with 2048 or more columns in the launch's batch, at 1 and 3-5 rows, and b otherwise. c is no longer picked.
• 6 to 16 rows are left to dsv41 metal: launches of 6 to 16 rows run their FP8 projections on a matrix tile (stacked on #1720) #2973's tile kernel. At those widths b's new internals are about level (FP8 sum 2-4% faster, wo_b at 16 rows 4.6% slower).

Measured on an M5 Ultra (512 GB) on release layer 2, the eight FP8 launches each timed alone (time-dsv41-metal-real, TF_DSV41_TIMING_FP8=1, mean of two balanced runs):

Rows	Before	After
1	247.4 µs (675 GB/s)	214.8 µs (778 GB/s)
4	325.5 µs	281.4 µs
5	359.8 µs	316.0 µs

• Engram's wkv (6144 x 25600) moves to kind d: 311-365 → 230-237 µs at one row.

• Served on the release weights with dsv41 metal: launches of 6 to 16 rows run their FP8 projections on a matrix tile (stacked on #1720) #2973 and dsv41 metal: decode experts decode an expert's weights once for all of its rows (stacked on #1720, #2973) #2983 underneath, 16 prompts of 512 tokens: draft:false 46.4 → 49.5 tok/s and greedy 82.2 → 84.3, with every reply's text equal to main's (48 of 48).

• Bits: release layers 2 and 21 give main's output and KV hashes at 1 and 4 rows, 20 of 20 concurrent launches equal, and the float64 real-layer checks pass (layer 2: 362 passes, layer 21: 266).

• Kernel contract, --only "dsv41_fp8_rows": 6 kernels pass gates 1-4, 4 controls red, and 288 invariance launches byte-equal. The probe's invariance loop now covers every kind.

• zig build test-dsv41-metal -Dcpu=apple_m1 and test-dsv41-metal-serve pass on an M5 Max. Lean: 0 new.

• Tried and dropped: kind c for K=1280, 2-4 adjacent groups a lane, 2-4 column blocks a threadgroup, X held in registers across columns, applying the scale once at the end of a chain (changes subnormal and overflow bits).

• Synthena (Ash's Agent)