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

Simplify dequantisation

This commit is contained in:
Ed Addario 2025-09-27 17:27:58 +01:00
parent 8a2c71f471
commit 3d75b14c0f
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 10 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
src/llama-quant.cpp
View File

@ -819,25 +819,16 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
// Dequantize into dequantized_buffer // Dequantize into dequantized_buffer
{ {
const ggml_type_traits * traits = ggml_get_type_traits(quant_type); const ggml_type_traits * traits = ggml_get_type_traits(quant_type);
if (traits && traits->to_float && quant_type != GGML_TYPE_F16 && quant_type != GGML_TYPE_BF16) { if (!traits || !traits->to_float) {
traits->to_float(quantized_buffer.data(), dequantized_buffer.data(), (int)(sample_rows * (size_t)n_per_row)); if (out_mse) { *out_mse = infinity; }
} else { if (out_proj) { *out_proj = 0.0; }
for (size_t r = 0; r < sample_rows; ++r) { return infinity;
const uint8_t * src = quantized_buffer.data() + r * row_sz; }
float * dst = dequantized_buffer.data() + r * (size_t)n_per_row;
if (quant_type == GGML_TYPE_F16) { for (size_t r = 0; r < sample_rows; ++r) {
ggml_fp16_to_fp32_row((const ggml_fp16_t *)src, dst, (int)n_per_row); const uint8_t * src = quantized_buffer.data() + r * row_sz;
} else if (quant_type == GGML_TYPE_BF16) { float * dst = dequantized_buffer.data() + r * (size_t)n_per_row;
ggml_bf16_to_fp32_row((const ggml_bf16_t *)src, dst, (int)n_per_row); traits->to_float(src, dst, (int)n_per_row);
} else {
if (!traits || !traits->to_float) {
if (out_mse) { *out_mse = infinity; }
if (out_proj) { *out_proj = 0.0; }
return infinity;
}
traits->to_float(src, dst, (int)n_per_row);
}
}
} }
} }