metal : optimize argmax and pool_1d (#30862)

* metal : optimize argmax and pool_1d

Add a two-pass partial reduction for large-row argmax and use cooperative threadgroup reductions for pool_1d. Allocate scratch space for the argmax partials through the Metal buffer allocator.

Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* tests : add pool_1d perf cases

Add representative POOL_1D max cases for the MoE down-projection amax shapes, reducing the whole intermediate dim.

Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* cont : minor fix
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2026-10-20 16:46:00 +03:00
1 parent 9df0b73c70
commit df11702136
10 files changed
+341 -45

No files matched your search

+11
View File
@@ -12199,6 +12199,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
}
}
// POOL_1D max over MoE down-projection amax shapes:
// amax is [1, n_expert, n_tokens] after reducing the whole intermediate dim.
for (int64_t n_ff : { 2048, 4096, 8192 }) {
for (int64_t n_expert : { 8, 16 }) {
for (int64_t n_tokens : { 1, 512 }) {
test_cases.emplace_back(new test_pool1d(
GGML_OP_POOL_MAX, GGML_TYPE_F32, {n_ff, n_expert, n_tokens, 1}, n_ff, n_ff, 0));
}
}
}
// LEAKY_RELU at FFN activation width, for direct comparison with RELU
for (int64_t n_tokens : {512, 2048}) {
test_cases.emplace_back(new test_leaky_relu(GGML_TYPE_F32, { 17408, n_tokens, 1, 1 }, 0.1f));