Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 9 additions & 2 deletions tools/ci/hrx_f16_narrowing.py
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,9 @@

This is the ratchet:

* every .loom body that narrows f32 to f16 (`f32 to f16` or `truncf`) is listed in
* every .loom body that narrows f32 to f16 -- by the `fptrunc` op, the scalar `truncf` spelling, or an
`f32 to f16` type line, which covers both the scalar and the `vector<Nxf32> to vector<Nxf16>` forms -- is
listed in
tools/ci/hrx_f16_narrowing.txt, and a body that appears without a line, or a line whose body is gone,
fails the check -- the two arms a ratchet needs (a list that can only grow cannot see a rename,
engine#1728's lesson);
Expand All @@ -45,7 +47,12 @@

CORPUS = pathlib.Path("third_party/llama.cpp/ggml/src/ggml-hrx/kernel-corpus/kernels")
BASELINE = pathlib.Path("tools/ci/hrx_f16_narrowing.txt")
NARROW = ("f32 to f16", "truncf")
# engine#2219: `truncf` and `f32 to f16` between them miss `vector.fptrunc %v : vector<4xf32> to vector<4xf16>`
# -- the vector spelling contains neither `truncf` (it is `fptrunc`) nor `f32 to f16` (the types are wrapped in
# `vector<Nx...>`). At pin 2b7960bae that form is 132 of the 233 fptrunc lines and 27 of the 52 narrowing bodies,
# so the ratchet was listing 25 of 52 and reporting "none new, none stale". `fptrunc` is LLVM's narrowing op and
# covers both spellings; the other markers stay because a hand-written or vendor file may use them without it.
NARROW = ("fptrunc", "truncf", "f32 to f16")
GUARD = ("scale_log2", "prescale")


Expand Down
27 changes: 27 additions & 0 deletions tools/ci/hrx_f16_narrowing.txt
Original file line number Diff line number Diff line change
Expand Up @@ -21,28 +21,55 @@
# perfectly correct. What the list buys is that a NEW narrowing body cannot appear without someone
# writing it down and saying what feeds it -- and that is the open question on engine#1962: which of
# these are protected by a producer-side prescale cannot be answered by reading them.
loom-libs/motifs/bf16_f16.loom
loom-libs/motifs/dequant.loom
loom-libs/motifs/dequant_1bit.loom
loom-libs/motifs/dequant_prism.loom
loom-libs/motifs/f32_f16.loom
loom-libs/motifs/llm_attention_qkv_matmul_postops.loom
loom-libs/motifs/mul_mat_f32_f32_wmma_core.loom
loom-libs/motifs/mul_mat_id_f16_f16_wmma_core.loom
loom-libs/motifs/mul_mat_id_f32_f32_wmma_core.loom
loom-libs/motifs/mul_mat_id_q4k_f16_wmma_projection.loom
loom-libs/motifs/mul_mat_id_q5k_f16_wmma_projection.loom
loom-libs/motifs/mul_mat_id_q6k_f16_wmma_projection.loom
loom-libs/motifs/mul_mat_id_swiglu_f16_f16_accumulate.loom
loom-libs/motifs/mul_mat_id_swiglu_f16_f16_wmma_body.loom
loom-libs/motifs/mul_mat_quantized_f16_prefill.loom
loom-libs/motifs/publish_f32.loom
loom-libs/motifs/q4_k_f16.loom
loom-libs/motifs/q6_k_f16.loom
loom-libs/motifs/q8_0_f16.loom
loom-libs/motifs/q8_1_f16.loom
loom-libs/motifs/quantize_q8_1_x4.loom
loom-libs/motifs/unary_f32_apply.loom
loom-libs/ops/copy_f32.loom
loom-libs/ops/flash_attention_decode_split_f32_f16_wmma.loom
loom-libs/ops/flash_attention_f32_f16_wmma.loom
loom-libs/ops/gated_delta_net_f32_wmma.loom
loom-libs/ops/llm_attention_k_matmul_rope_set_rows_decode_f32_f32.loom
loom-libs/ops/llm_attention_v_matmul_set_rows_decode_f32_f32.loom
loom-libs/ops/mul_mat_id_swiglu_f32_f32_wmma.loom
loom-libs/ops/mul_mat_q6_k_packed_f16_wmma.loom
loom-libs/ops/mul_mat_swiglu_f32_f32_wmma.loom
loom-libs/ops/mul_mat_swiglu_symmetric_i4_wmma.loom
loom-libs/ops/mul_mat_symmetric_i4_wmma.loom
loom-libs/ops/mul_mat_symmetric_i8_wmma.loom
loom-libs/ops/rmsnorm_binary_f32.loom
loom-libs/ops/rope_set_rows_f32.loom
loom-libs/ops/set_rows.loom
loom-libs/ops/small_rows_f32.loom
qwen_moe/ggml/linear_q6k_q8_1_x4.loom
qwen_moe/ggml/quantize_q8_1_x4.loom
qwen_moe/qwen3_moe/attention_postprocess_f32_f16.loom
qwen_moe/qwen3_moe/attention_prepare_quantized.loom
qwen_moe/qwen3_moe/dense_linear_quantized_f16_wmma.loom
qwen_moe/qwen3_moe/flash_attention_decode_f32_f16_wmma.loom
qwen_moe/qwen3_moe/flash_attention_decode_q128_f32_f16_wmma.loom
qwen_moe/qwen3_moe/flash_attention_decode_split_f32_f16_wmma.loom
qwen_moe/qwen3_moe/flash_attention_decode_split_next_q8_test.loom
qwen_moe/qwen3_moe/flash_attention_f32_f16_wmma.loom
qwen_moe/qwen3_moe/routed_down_quantized_f16_wmma.loom
qwen_moe/qwen3_moe/routed_down_weighted_reduce_next_rmsnorm_q8_1_x4.loom
qwen_moe/qwen3_moe/routed_gate_up_swiglu_q4k.loom
qwen_moe/qwen3_moe/routed_linear_q4k_f16_wmma.loom
Loading