vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : fix BLAS with unsupported types (#9775)

Browse source 
* ggml : do not use BLAS with types without to_float

* ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies

* ggml : rename ggml_internal_get_type_traits -> ggml_get_type_traits

it's not really internal if everybody uses it
This commit is contained in:
Diego Devesa 2024-10-08 14:21:43 +02:00 • committed by GitHub
parent 458367a906
commit dca1d4b58a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 75 additions and 74 deletions
Download patch file Download diff file Expand all files Collapse all files

9
src/llama.cpp
View file

@ -17872,10 +17872,9 @@ static void llama_tensor_dequantize_internal(
}
float * f32_output = (float *) output.data();
ggml_type_traits_t qtype;
const ggml_type_traits * qtype = ggml_get_type_traits(tensor->type);
if (ggml_is_quantized(tensor->type)) {
qtype = ggml_internal_get_type_traits(tensor->type);
if (qtype.to_float == NULL) {
if (qtype->to_float == NULL) {
throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
}
} else if (tensor->type != GGML_TYPE_F16 &&
@ -17889,7 +17888,7 @@ static void llama_tensor_dequantize_internal(
} else if (tensor->type == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
} else if (ggml_is_quantized(tensor->type)) {
qtype.to_float(tensor->data, f32_output, nelements);
qtype->to_float(tensor->data, f32_output, nelements);
} else {
GGML_ABORT("fatal error"); // unreachable
}
@ -17925,7 +17924,7 @@ static void llama_tensor_dequantize_internal(
} else if (typ == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
} else {
qtype.to_float(inbuf, outbuf, nels);
qtype->to_float(inbuf, outbuf, nels);
}
};
workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);