metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT (#30100)
* metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT ggml_metal_op_mul_mat_mma picks the residual of a fused MUL_MAT+ADD as "the ADD operand whose op is not MUL_MAT". When both operands of the ADD are mat-mul outputs (x = W1 @ u + W2 @ v), that test is true for both, so the residual resolves to the fused mat-mul's own, never-written output and the kernel adds whatever that buffer holds. The fusion check (ggml_metal_mul_mat_add_operand) already selects the operand by identity; make the encoder do the same. Clef decision models hit this in their head (proj_option_context @ ctx + proj_option_lexical @ lex, 9 option rows): on Metal, /v1/systemone probabilities collapse toward uniform (billing 0.28 where the CPU backend gives 0.977, Cloudflare_clef-flash Q8_0), deterministic per memory layout, correct with GGML_METAL_FUSION_DISABLE=1. Not a quantization issue: the same file is right on CPU. Add a MUL_MAT_ADD mode to test-backend-ops where the residual is a second mat-mul; on Metal it fails 27 of 28 cases before this change (the one pass is f16 n=2, under the MMA row threshold, so nothing fuses). * Update tests/test-backend-ops.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
1 parent
b9acf138a1
commit
7e8324f5fe
2 files changed
+11
-2
No files matched your search
@@ -2476,7 +2476,8 @@ static int ggml_metal_op_mul_mat_mma(ggml_metal_op_t ctx, int idx) {
|
||||
|
||||
if (fuse_add) {
|
||||
dst = ctx->node(idx + n_fuse - 1);
|
||||
res = dst->src[0]->op == GGML_OP_MUL_MAT ? dst->src[1] : dst->src[0];
|
||||
// the residual is the other operand of the ADD, by identity: it can itself be a MUL_MAT output
|
||||
res = dst->src[0] == op ? dst->src[1] : dst->src[0];
|
||||
}
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_mul_mv_mma_auto(lib, op, fuse_add);
|
||||
|
||||
@@ -7233,6 +7233,7 @@ enum mul_mat_add_mode {
|
||||
MUL_MAT_ADD_ROW, // mm + a one-row res, which a same-shape fusion must leave alone
|
||||
MUL_MAT_ADD_RES_INPLACE, // res += mm
|
||||
MUL_MAT_ADD_B_INPLACE, // b += mm: the sum overwrites the mat-mul input (m == k)
|
||||
MUL_MAT_ADD_MM_MM, // mm2 + mm: the residual is itself a mat-mul output, so both operands are MUL_MAT
|
||||
};
|
||||
|
||||
static std::string var_to_str(mul_mat_add_mode mode) {
|
||||
@@ -7242,6 +7243,7 @@ static std::string var_to_str(mul_mat_add_mode mode) {
|
||||
case MUL_MAT_ADD_ROW: return "mm+row";
|
||||
case MUL_MAT_ADD_RES_INPLACE: return "res+=mm";
|
||||
case MUL_MAT_ADD_B_INPLACE: return "b+=mm";
|
||||
case MUL_MAT_ADD_MM_MM: return "mm2+mm";
|
||||
}
|
||||
return "unknown";
|
||||
}
|
||||
@@ -7287,6 +7289,12 @@ struct test_mul_mat_add : public test_case {
|
||||
case MUL_MAT_ADD_RES_MM: out = ggml_add(ctx, res, mm); break;
|
||||
case MUL_MAT_ADD_RES_INPLACE: out = ggml_add_inplace(ctx, res, mm); break;
|
||||
case MUL_MAT_ADD_B_INPLACE: out = ggml_add_inplace(ctx, b, mm); break;
|
||||
case MUL_MAT_ADD_MM_MM:
|
||||
{
|
||||
ggml_tensor * a2 = ggml_new_tensor_2d(ctx, type_a, k, m);
|
||||
ggml_tensor * mm2 = ggml_mul_mat(ctx, a2, b);
|
||||
out = ggml_add(ctx, mm2, mm);
|
||||
} break;
|
||||
}
|
||||
ggml_set_name(out, "out");
|
||||
|
||||
@@ -10409,7 +10417,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
|
||||
}
|
||||
}
|
||||
for (ggml_type type_a : {GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q5_K, GGML_TYPE_F16}) {
|
||||
for (mul_mat_add_mode mode : {MUL_MAT_ADD_MM_RES, MUL_MAT_ADD_RES_MM, MUL_MAT_ADD_ROW, MUL_MAT_ADD_RES_INPLACE}) {
|
||||
for (mul_mat_add_mode mode : {MUL_MAT_ADD_MM_RES, MUL_MAT_ADD_RES_MM, MUL_MAT_ADD_ROW, MUL_MAT_ADD_RES_INPLACE, MUL_MAT_ADD_MM_MM}) {
|
||||
test_cases.emplace_back(new test_mul_mat_add(type_a, 1000, n, 1024, mode));
|
||||
}
|
||||
test_cases.emplace_back(new test_mul_mat_add(type_a, 2048, n, 2048, MUL_MAT_ADD_B_INPLACE));
|
||||
|
||||
Reference in new issue
Block a user