happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Fix llama-3-8b and phi3-mini q4_0 NPU

This commit is contained in:
Yu, Zijun 2025-10-14 14:51:42 +08:00 committed by Mustafa Cavus
parent 299f4923bb
commit 2d2f00a41f
1 changed files with 17 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-openvino/ggml-quants.cpp
View File

@ -414,6 +414,13 @@ std::shared_ptr<ov::Node> requantize(const ggml_tensor* tensor, ExtraQuantType r
std::shared_ptr<ov::Node> weight_node; std::shared_ptr<ov::Node> weight_node;
ov::Shape node_shape = {(uint64_t) (tensor->ne[1]), (uint64_t) (tensor->ne[0])}; ov::Shape node_shape = {(uint64_t) (tensor->ne[1]), (uint64_t) (tensor->ne[0])};
// FIXME hardcoded workaround to fix the case where token_emb.weight is q4_0 (instead of q6_k)
// (In some q4_0 models which use two different weight for token_emb and output, token_emb is q4_0)
std::string device = getenv("GGML_OPENVINO_DEVICE") ? getenv("GGML_OPENVINO_DEVICE") : "";
if (device == "NPU" && std::string(tensor->name) == "token_embd.weight") {
requant_type = ExtraQuantType::F16;
}
if (requant_type == ExtraQuantType::F16) { if (requant_type == ExtraQuantType::F16) {
ov::Tensor weights(ov::element::f16, node_shape); ov::Tensor weights(ov::element::f16, node_shape);
ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), ggml_nelements(tensor)); ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), ggml_nelements(tensor));
@ -473,7 +480,16 @@ void quantize_q4_0(const float* x, ov::Tensor& weights_arr, ov::Tensor& scales_a
} }
const float d = max / -8; const float d = max / -8;
const float id = d ? 1.0f / d : 0.0f;
if (d == 0) {
scales[i] = ov::float16(1.0f);
biases[i] = ov::float16(-8.0f);
uint8_t zp = 8;
memset(weights + i * qk / 2, zp | (zp << 4), qk / 2);
continue;
}
const float id = 1.0f / d;
scales[i] = ov::float16(d); scales[i] = ov::float16(d);
biases[i] = ov::float16(-8.f * d); biases[i] = ov::float16(-8.f * d);