Strata

metal: the IQ2_S expert kernel, tiled - 515.7 GB/s effective, 89 % of the streaming ceiling

2026-10-23 18:51:05 +02:00

alu_bench.cpp

metal: dense GEMV for every pack format, prepared activations, window-size specialisation

2026-10-12 11:17:02 +02:00

arena_budget_test.cpp

metal: a hard cap on the engine's whole resident footprint, and no engine beside a repack (the two M5 Max watchdog panics)

2026-10-13 12:47:21 +02:00

coop_layout.cpp

metal: prompt experts and QSA attention on matrix-unit register operands (v3), prompt_chain defaults to them

2026-10-12 14:54:56 +02:00

coop_probe.h

metal: mm_gu3 paired-run decode without the runtime layout check (Q2_0 gate/up 4.9 -> 3.5 ms a layer); SLC prefetch probe

2026-10-12 17:23:39 +02:00

dense_mm_parity.cpp

metal: the Unsloth UD-Q4/5/6_K_XL formats - K-quant / Q8_0 / Q5_1 routed experts, F32 GEMV + router, Q8_0 HC read with F32 inject

2026-10-12 22:56:44 +02:00

df2_bench.cpp

metal: df2_mma8 (draft head Q8_0 rows on the matrix units), stubs_df2.cpp from the first Metal port

2026-11-01 23:11:58 +01:00

ds4_dq_parity.cpp

dsv41 Metal: router NaN guard in the MSL route kernels, and a dense Q4_K dequantizer; tests for both

2026-11-01 09:49:17 +01:00

ds4_route_guard_parity.cpp

DeepSeek engine: an expert profile of another model's size is not read; the router tests run with 192 experts and poisoned padding (CUDA, Metal)

2026-11-01 19:18:48 +01:00

dsu_job.metal

Metal Q6_K dense matvec: unaligned block loads (210-byte blocks), found by dsu_parity; test fix

2026-10-30 22:36:12 +01:00

dsu_parity.cpp

Metal: Q6_K multi-column dense matvec (ds4_q6k_mvn)

2026-10-30 22:36:13 +01:00

dsu_ref_vs_ggml.cpp

Metal tests for the Unsloth set kernels: IQ1_S matvec and tile loader, Q6_K dequant and matvec, reference checked against ggml

2026-10-30 22:36:12 +01:00

expert_parity.cpp

Metal: Q5_0 down experts, decode (moe.metal, moeu) and prompt (moe_mm v1/v2/v3), host tiles, honest gu_qk/d_qk; parity tests take Q5_0

2026-11-05 07:14:30 +01:00

fuse_parity.cpp

Metal: 256 routed experts (the expert-pruned Coder, #4706 ): the routers, the prompt chain and the SSD stream take the model's expert count

2026-10-31 20:54:38 +01:00

gdn_chunk_bench.cpp

Metal: GDN chunked v2 (measured, slower), M5 GEMM tiles + fused GDN input projections (bit-identical), layer bench

2026-10-30 01:42:07 +01:00

gdn_layer_bench.cpp

Metal: tensor-op chunk-state probe, layer bench with before/after, A/B driver

2026-10-30 01:50:15 +01:00

gdn_parity.cpp

Metal: gdn_step fits its launch size, every dispatch checks maxTotalThreadsPerThreadgroup, qsa_prompt compiles on macOS 26.3 (#2323 )

2026-10-13 23:21:28 +02:00

gdn_prompt_parity.cpp

metal: chunked GDN scan (WY form, chunks of 32) - correct, slower in scalar fp32 than gdn_scan_r; kept, not used

2026-10-12 17:32:53 +02:00

gdn_tprobe.cpp

Metal: tensor-op chunk-state probe, layer bench with before/after, A/B driver

2026-10-30 01:50:15 +01:00

gemv_parity.cpp

Metal: opt-in fp16 GEMVs and Q2_0 experts in decode windows (STRATA_METAL_F16=1), from PR 4807 by rasa04

2026-11-01 00:39:53 +01:00

gr2_parity.cpp

metal: GDN step writes the prepared activations of its output; token_chain --fuse / --fold-comb

2026-10-12 12:08:33 +02:00

gr_parity.cpp

metal: the Unsloth UD-Q4/5/6_K_XL formats - K-quant / Q8_0 / Q5_1 routed experts, F32 GEMV + router, Q8_0 HC read with F32 inject

2026-10-12 22:56:44 +02:00

gridbar_bench.cpp

metal: persistent-kernel probe (grid barrier between phases): correct only with device-scope fences, not faster

2026-10-12 12:30:33 +02:00

gyro_parity.sh

metal/tests/gyro_parity.sh: the Gyro kernel checks in one script

2026-10-20 18:40:57 +02:00

mm_bench.cpp

metal: Q8_0 HC read the default in token_chain (STRATA_HC_Q8=0 = exact BF16); prompt chain, GDN prompt scan, fused HC prompt read

2026-10-12 13:33:12 +02:00

mm_i8_bench.cpp

Metal prompt: host rows beside the GPU (embedding + PLE rows of the next chunk on a thread, parallel PLE gather on Macs), qsa_merge_h (merge writes the fp16 operand), ragged expert tiles (mm_gu6/mm_down6/mm_scan6, STRATA_PROMPT_V6), STRATA_PROMPT_KPROF per-dispatch profile, matmul2d probes

2026-11-05 20:36:54 +01:00

mma_parity.cpp

metal: simdgroup-matrix GEMV for windows (exact, measured slower; kept as a reference path)

2026-10-12 12:15:03 +02:00

moe_mm_parity.cpp

Metal prompt: host rows beside the GPU (embedding + PLE rows of the next chunk on a thread, parallel PLE gather on Macs), qsa_merge_h (merge writes the fp16 operand), ragged expert tiles (mm_gu6/mm_down6/mm_scan6, STRATA_PROMPT_V6), STRATA_PROMPT_KPROF per-dispatch profile, matmul2d probes

2026-11-05 20:36:54 +01:00

pair_parity.cpp

metal: gemv_pair (shared expert gate + up + SwiGLU in one dispatch), raw-activation Q2_0 GEMV; pair measured slower

2026-10-12 12:33:59 +02:00

prompt_chain.cpp

Metal: 256 routed experts (the expert-pruned Coder, #4706 ): the routers, the prompt chain and the SSD stream take the model's expert count

2026-10-31 20:54:38 +01:00

prompt_parity.cpp

Metal: 256 routed experts (the expert-pruned Coder, #4706 ): the routers, the prompt chain and the SSD stream take the model's expert count

2026-10-31 20:54:38 +01:00

qsa_parity.cpp

metal: K-quant / Q5_1 paired-run prompt decoders (W8) and paired mm_down3; qsa_prep M-RoPE (function constant 24)

2026-10-12 23:10:47 +02:00

qsa_prompt_parity.cpp

metal: QSA prompt attention correct on the M4 (K = 32 runs, threadgroup-staged store, fresh per-chunk accumulator)

2026-10-12 16:57:55 +02:00

ref_dump.cpp

metal: standalone Metal kernel tree - runtime face, bandwidth/dispatch probes, grouped routed-expert kernels

2026-10-12 10:42:59 +02:00

route_parity.cpp

metal: whole-model decode step on real GGUF weights (token_chain), dependency-graph encoding

2026-10-12 11:48:27 +02:00

rt_bench.cpp

metal: dense GEMV for every pack format, prepared activations, window-size specialisation

2026-10-12 11:17:02 +02:00

runtime_hello.cpp

metal: the Metal runtime behind a CUDA-name compat layer (the HIP model), so the host files build for Metal unchanged

2026-10-12 11:53:32 +02:00

sample_chain_parity.cpp

metal: the sampler chain on the GPU (sample.metal samp_part + samp_pick: bias, penalties, top_k, top_p, min_p, temperature, Philox draw at (seed, position)) and sample_chain_parity (greedy == argmax, 1800 rows against the CPU chain with 0 mismatches, determinism, chi-square)

2026-11-02 05:52:27 +01:00

sample_parity.cpp

metal: QSA decode (prep, block scores, top-k, split-K attention, gated merge) and the greedy sampler

2026-10-12 11:38:40 +02:00

serve_probe.py

metal: serve_probe.py, the owner round 9 short/long prompt probe on the default config

2026-10-22 13:05:58 +02:00

token_chain.cpp

Metal: token_chain CHAIN_GDNGRID timing probe (gdn_step grid size)

2026-10-31 15:39:20 +01:00

wrap_split_test.cpp

metal: wrap_split_test: the note's newline

2026-10-15 07:02:42 +02:00