Skip to content

llama: disable fused up-gate when an offload backend lacks GGML_OP_FUSED_UP_GATE (fixes silent 3.5x Metal slowdown) #2133

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

Open
hchengit wants to merge 1 commit into
base: main
Choose a base branch
from
Open

llama: disable fused up-gate when an offload backend lacks GGML_OP_FUSED_UP_GATE (fixes silent 3.5x Metal slowdown) #2133

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
33 changes: 33 additions & 0 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7972,6 +7972,39 @@ struct llama_context * llama_init_from_model(
}
ctx->backends.push_back(ctx->backend_cpu);

// Fused up-gate nodes are emitted into the graph whenever the option is
// enabled, but not every backend implements GGML_OP_FUSED_UP_GATE (e.g. the
// Metal backend does not). The scheduler silently assigns those nodes to the
// CPU backend, forcing a device<->host round trip on every FFN layer when the
// rest of the graph is offloaded. Probe the offload backends and fall back to
// the decomposed ops (MUL_MAT + FUSED_MUL_UNARY) up front instead.
if (cparams.fused_up_gate) {
struct ggml_init_params probe_init = {
/*.mem_size =*/ 8*ggml_tensor_overhead(),
/*.mem_buffer =*/ nullptr,
/*.no_alloc =*/ true,
};
ggml_context * probe_ctx = ggml_init(probe_init);
ggml_tensor * up = ggml_new_tensor_2d(probe_ctx, GGML_TYPE_Q4_0, 256, 16);
ggml_tensor * gate = ggml_new_tensor_2d(probe_ctx, GGML_TYPE_Q4_0, 256, 16);
ggml_tensor * b = ggml_new_tensor_2d(probe_ctx, GGML_TYPE_F32, 256, 1);
ggml_tensor * probe = ggml_fused_up_gate(probe_ctx, up, gate, b, GGML_UNARY_OP_SILU);
GGML_ASSERT(probe->op == GGML_OP_FUSED_UP_GATE);
for (auto * backend : ctx->backends) {
if (backend == ctx->backend_cpu) {
continue;
}
if (!ggml_backend_supports_op(backend, probe)) {
LLAMA_LOG_WARN("%s: backend %s does not implement %s - disabling fused up-gate "
"(decomposing into mul_mat + fused_mul_unary)\n",
__func__, ggml_backend_name(backend), ggml_op_name(probe->op));
cparams.fused_up_gate = false;
break;
}
}
ggml_free(probe_ctx);
}

if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, params.idx_type_k, kv_size, cparams.offload_kqv,
params.type_k_first, params.type_k_last, params.type_v_first, params.type_v_last,
params.n_k_first, params.n_k_last, params.n_v_first, params.n_v_last)) {
Expand Down