Skip to content

CUDA: honor the F32 precision Inkling asks for on the banded attention op #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
LeoBorcherding wants to merge 1 commit into
base: night-shift-tester/audit-pin-inkling-534c321
Choose a base branch
from
Draft

CUDA: honor the F32 precision Inkling asks for on the banded attention op #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 6 additions & 0 deletions ggml/src/ggml-cuda/fattn.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -543,6 +543,12 @@ bool ggml_cuda_flash_attn_ext_banded_use_mma(const int device, const ggml_tensor
return false;
}

// the MMA instance accumulates V*softmax in fp16 on NVIDIA, so an F32 request takes the FP32 kernel
const ggml_prec prec = ggml_flash_attn_ext_get_prec(dst);
if (prec != GGML_PREC_UNDEFINED && prec <= GGML_PREC_F32) {
return false;
}

const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
Expand Down
20 changes: 20 additions & 0 deletions tests/test-backend-ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8596,6 +8596,25 @@ struct test_flash_attn_ext_banded : public test_case {
}
};

// nonzero-mean V at long n_kv: an fp16 VKQ accumulator drops increments once the running sum is large,
// which zero-mean V can't show. GGML_PREC_F32 is requested, so hold it to the F32 tolerance.
struct test_flash_attn_ext_banded_v_offset : public test_flash_attn_ext_banded {
using test_flash_attn_ext_banded::test_flash_attn_ext_banded;

std::string vars() override {
return test_flash_attn_ext_banded::vars() + ",v=U(900,1100)";
}

double max_nmse_err() override {
return 2e-6;
}

void initialize_tensors(ggml_context * ctx) override {
test_flash_attn_ext_banded::initialize_tensors(ctx);
init_tensor_uniform(ggml_get_tensor(ctx, "v"), 900.0f, 1100.0f);
}
};

// large Q values, so the online softmax has to rescale the partial results
struct test_flash_attn_ext_large_logits : public test_flash_attn_ext {
static constexpr int q_range = 20;
Expand Down Expand Up @@ -11897,6 +11916,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_flash_attn_ext_banded(128, 8, 1, 512, 17024, 1024, 1, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_flash_attn_ext_banded(128, 8, 1, 1, 17024, 1024, 1, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_flash_attn_ext_banded(128, 8, 1, 512, 32768, 1024, 1, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_flash_attn_ext_banded_v_offset(128, 8, 1, 512, 131072, 1024, 1, GGML_TYPE_F16, GGML_TYPE_F32));

test_cases.emplace_back(new test_flash_attn_ext(256, 256, 4, {6, 1}, 512, 512, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
test_cases.emplace_back(new test_flash_attn_ext(256, 256, 4, {6, 1}, 4096, 64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
Expand Down
Loading