llama: move moe-cache code out of llama-graph (#32157)

This commit is contained in:
Aman Gupta authored and GitHub committed 2026-11-13 18:48:29 +08:00
1 parent 2d8a325029
commit 951d465f5c
4 files changed
+121 -112

No files matched your search

+9 -74
View File
@@ -2425,12 +2425,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
// compute the expert input before the slot lookup, so that the scheduler prefetches it for the CPU part (the router can read a different tensor)
ggml_build_forward_expand(gf, cur);
// the experts of host-resident layers may be read from the MoE cache, and part of the misses computed from the host experts
// the CPU part does not support weights before the FFN, LoRAs or rotated experts
const std::array<ggml_tensor *, 4> exps = { up_exps, gate_exps, down_exps, gate_up_exps };
const bool can_cpu_share = !weight_before_ffn && loras->empty() &&
std::none_of(exps.begin(), exps.end(), [&](ggml_tensor * w) { return w != nullptr && hdmd != nullptr && hdmd->rot.count(w) > 0; });
const llm_graph_moe_cache_slots cache = build_moe_cache_slots(selected_experts, up_exps, gate_exps, down_exps, gate_up_exps, can_cpu_share, il);
// the experts of host-resident layers may be read from the MoE cache (see llama_moe_cache::build_lookup)
const llama_moe_cache_lookup cache = moe_cache == nullptr ? llama_moe_cache_lookup{} :
moe_cache->build_lookup(*this, { il, up_exps, gate_exps, down_exps, gate_up_exps, weight_before_ffn }, selected_experts);
cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
@@ -2603,14 +2600,12 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
return experts;
};
ggml_tensor * experts = build_experts(cur, selected_experts, cache.slots, true);
if (cache.cpu_ids != nullptr) {
// the cached experts and the misses computed from the host experts, each reads the others as zero
// the cached experts come first in the graph, so that they are computed while the misses are
ggml_build_forward_expand(gf, experts);
experts = ggml_add(ctx0, experts, build_experts(cur, selected_experts, cache.cpu_ids, false));
cb(experts, "ffn_moe_down_merged", il);
}
ggml_tensor * experts = cache.slots == nullptr ?
build_experts(cur, selected_experts, nullptr, true) :
moe_cache->build_experts(*this, cache, cur, selected_experts,
[](ggml_tensor * cur, ggml_tensor * ids, ggml_tensor * mm_ids, bool cached, void * user_data) {
return (*static_cast<decltype(build_experts) *>(user_data))(cur, ids, mm_ids, cached);
}, &build_experts);
if (down_exps_s) {
cb(experts, "ffn_moe_down_scaled", il);
@@ -2669,66 +2664,6 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
return moe_out;
}
llm_graph_moe_cache_slots llm_graph_context::build_moe_cache_slots(
ggml_tensor * selected_experts,
ggml_tensor * up_exps,
ggml_tensor * gate_exps,
ggml_tensor * down_exps,
ggml_tensor * gate_up_exps,
bool cpu_part,
int il) const {
if (moe_cache == nullptr) {
return { nullptr, nullptr };
}
ggml_tensor * cpu_map = cpu_part ? moe_cache->get_cpu_map(il, selected_experts->ne[1]) : nullptr;
ggml_tensor * slot_map = moe_cache->get_slot_map(il, selected_experts->ne[1], selected_experts->ne[0], cpu_map != nullptr);
if (slot_map == nullptr) {
return { nullptr, nullptr };
}
for (ggml_tensor * w : { up_exps, gate_exps, down_exps, gate_up_exps }) {
if (w == nullptr) {
continue;
}
ggml_tensor * experts = moe_cache->get_experts(w);
if (experts == nullptr) {
return { nullptr, nullptr };
}
// the device of the cache must compute the matmul, else the scheduler copies the cache to another backend
ggml_tensor * b = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, experts->ne[0], w == down_exps ? selected_experts->ne[0] : 1, selected_experts->ne[1]);
if (!ggml_backend_supports_op(moe_cache->backend(il), ggml_mul_mat_id(ctx0, experts, b, selected_experts))) {
return { nullptr, nullptr };
}
}
ggml_tensor * ids = selected_experts;
if (!ggml_is_contiguous(ids)) {
ids = ggml_cont(ctx0, ids);
}
ids = ggml_reshape_1d(ctx0, ids, ggml_nelements(ids));
// the slot map is a host weight, so the scheduler starts a new split here and copies it with the copy callback
// the callback reads the selected experts, uploads the missing ones and updates the slot map
ggml_tensor * slots = ggml_get_rows(ctx0, slot_map, ids); // [1, n_expert_used*n_tokens]
if (!ggml_backend_supports_op(moe_cache->backend(il), slots)) {
return { nullptr, nullptr };
}
ggml_backend_sched_set_tensor_backend(sched, slots, moe_cache->backend(il));
cb(slots, "ffn_moe_slots", il);
// add the lookup now: the transform of a Hadamard-folded expert reads host weights and would start its split first
ggml_build_forward_expand(gf, slots);
// the misses computed from the host experts, read after the lookup has picked them
ggml_tensor * cpu_ids = nullptr;
if (cpu_map != nullptr) {
cpu_ids = ggml_get_rows(ctx0, cpu_map, ids); // [1, n_expert_used*n_tokens]
cpu_ids = ggml_reshape_2d(ctx0, cpu_ids, selected_experts->ne[0], selected_experts->ne[1]);
}
return { ggml_reshape_2d(ctx0, slots, selected_experts->ne[0], selected_experts->ne[1]), cpu_ids };
}
// input embeddings with optional lora
ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd, float tok_scale) const {
const int64_t n_embd_inp = hparams.n_embd_inp();