Niko1221 bfbe7fab2b MMQ parity screen: skip the trellis types when the ggml has no host traits for them (it read past the type table on Windows)

2026-11-05 08:06:52 +01:00

..

expert_cache_staging.cpp

AMD HIP backend for gfx1100 (PR #121, squashed onto its 0.1.21 base)

2026-09-29 14:33:45 +02:00

gdn_chunked.cpp

GDN: the chunked WMMA recurrence on gfx12 (opt-in STRATA_GDN_CHUNKED=1), gfx12 fragment layout

2026-10-30 00:59:13 +01:00

gdn_rec_head.cpp

prefill: STRATA_GDN_CHUNKED=1 on gfx11, the prompt's DeltaNet recurrence in 16-token chunks on the matrix cores (opt-in)

2026-10-14 09:59:02 +02:00

gdn_stage_bench.cpp

GDN gfx12: conv + SiLU + q/k norm written as FP16 and read by the chunked kernels (one entry, STRATA_GDN_FUSE=0 off); ablation hooks removed

2026-10-30 01:51:05 +01:00

gemm_f16_io_parity.cpp

tests: gemm_f16_io_parity builds its BF16 weights by hand and refuses an all-zero reference (#835)

2026-10-05 18:34:29 +02:00

handoff.cpp

AMD HIP backend for gfx1100 (PR #121, squashed onto its 0.1.21 base)

2026-09-29 14:33:45 +02:00

intrinsics.cpp

hip: add experimental gfx1012 support and legacy BLAS compatibility

2026-10-03 23:33:45 +02:00

mapped_alias.cpp

HIP on Windows: the hipBLAS sticky error (Windows only) and the mapped embedding fallback (#247, #325)

2026-10-01 20:20:09 +02:00

native_qsa_score.cpp

AMD HIP backend for gfx1100 (PR #121, squashed onto its 0.1.21 base)

2026-09-29 14:33:45 +02:00

ple_iq4.cpp

AMD HIP backend for gfx1100 (PR #121, squashed onto its 0.1.21 base)

2026-09-29 14:33:45 +02:00

prefill_gemm.cpp

tests: hip_prefill_gemm crosses an RDNA2 SGEMM slice boundary (T 3300, K 10240; FP16 beta 1 with a padded ldy, and BF16)

2026-10-14 07:11:04 +02:00

prefill_gr_write_bench.cpp

prefill timing: the hyper-connection write-back on its own line (it was charged to the phase before it, e.g. the wait for the first chunk's PLE rows); hip_prefill_gr_write_bench

2026-10-18 04:57:28 +02:00

prefill_hcd_exact_parity.cpp

prefill: STRATA_HCD_EXACT says once when it cannot take the kernel (no tuning table, or a solution other than 1176 / 1177), and hip_prefill_hcd_exact_parity checks the kernel against hipBLASLt

2026-10-05 11:27:31 +02:00

prefill_hipblaslt_gemm.cpp

hip: address review of the gfx1201 table: current docs, smoke-test wording

2026-10-01 17:57:41 +02:00

prefill_mmq_parity.cpp

MMQ parity screen: skip the trellis types when the ggml has no host traits for them (it read past the type table on Windows)

2026-11-05 08:06:52 +01:00

prefill_native_batch.cpp

tests/hip: prefill_native_batch passes the indexer a RopeScaling (the #84 signature; plain rope over base 1e4 as before) - a HIP-only test the CUDA build never compiled

2026-09-30 19:18:41 +02:00

prefill_rocblas_gemm.cpp

hip: a rocBLAS solution table for the FP16 prompt GEMMs on gfx103x (the GDN projection 7x below 1,152 tokens; short prompts -10 to -14%)

2026-10-14 07:12:35 +02:00

prefill_wmma_gemm_parity.cpp

tests: hip_prefill_wmma_gemm_parity skips on a card without matrix cores instead of failing (#1148)

2026-10-07 03:02:56 +02:00

q2_zero.cpp

hip: preserve Q2_0 signed zero on gfx1012 with legacy HIP

2026-10-03 23:33:46 +02:00

router_fast.cpp

HIP (S6): router_top10 without its serial parts on AMD, bitwise the same (default on HIP; STRATA_HIP_ROUTER_OLD=1: old)

2026-10-01 07:24:19 +02:00