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

imatrix : fix conversion problems

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-09-08 10:04:01 -04:00
parent 3de9300c37
commit c8ab6a3ba3
3 changed files with 14 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

8
convert_legacy_imatrix_to_gguf.py
View file

@ -64,10 +64,11 @@ class IMatrixReader:
data = self._get(np.float32, nval) data = self._get(np.float32, nval)
assert name not in self.entries, f"duplicated name: {name!r}" assert name not in self.entries, f"duplicated name: {name!r}"
self.entries[name] = IMatrixEntry(data, np.array([ncall * self.chunk_size], dtype=np.float32)) self.entries[name] = IMatrixEntry(data * np.float32(self.chunk_size), np.array([ncall * self.chunk_size], dtype=np.float32))
self.chunk_count = self._get(np.int32).item() self.chunk_count = self._get(np.int32).item()
self.dataset = self._get(np.uint8, self._get(np.int32).item()).tobytes().decode("utf-8") dataset_len = self._get(np.int32).item()
self.dataset = self._get(np.uint8, dataset_len).tobytes().decode("utf-8")
def to_writer(self, outfile: Path) -> IMatrixWriter: def to_writer(self, outfile: Path) -> IMatrixWriter:
writer = IMatrixWriter(path=outfile, arch="") writer = IMatrixWriter(path=outfile, arch="")
@ -110,6 +111,9 @@ if __name__ == "__main__":
input_file: Path = args.imatrix input_file: Path = args.imatrix
if input_file.suffix != ".gguf": if input_file.suffix != ".gguf":
args.outfile = input_file.with_suffix(".gguf") args.outfile = input_file.with_suffix(".gguf")
if args.outfile.exists():
logger.error(f"default file exists, specify with --outfile to overwrite: {args.outfile}")
exit(1)
writer = IMatrixReader(args.imatrix).to_writer(args.outfile) writer = IMatrixReader(args.imatrix).to_writer(args.outfile)

2
examples/imatrix/imatrix.cpp
View file

@ -31,7 +31,7 @@ static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size"; static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size";
struct Stats { struct Stats {
std::vector<double> values; std::vector<float> values;
std::vector<int64_t> counts; std::vector<int64_t> counts;
}; };

11
examples/quantize/quantize.cpp
View file

@ -132,6 +132,7 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
}; };
struct gguf_context * ctx_gguf = gguf_init_from_file(imatrix_file.c_str(), meta_gguf_params); struct gguf_context * ctx_gguf = gguf_init_from_file(imatrix_file.c_str(), meta_gguf_params);
if (!ctx_gguf) { if (!ctx_gguf) {
fprintf(stderr, "%s: if this is an older imatrix file, make sure to convert it to the GGUF-based imatrix format\n", __func__);
exit(1); exit(1);
} }
const int32_t n_entries = gguf_get_n_tensors(ctx_gguf); const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
@ -189,9 +190,9 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
e.resize(ggml_nelements(sums)); e.resize(ggml_nelements(sums));
float max_count = 0.0f; float max_count = 0.0f;
for (int64_t j = 0; j < ne1; ++j) { for (int64_t j = 0; j < ne1; ++j) {
const float count = ((const float *) counts->data)[ne1]; const float count = ((const float *) counts->data)[j];
for (int64_t i = 0; i < ne0; ++i) { for (int64_t i = 0; i < ne0; ++i) {
e[ne1*ne0 + ne0] = ((const float *) sums->data)[ne1*ne0 + ne0] / count; e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count;
} }
if (count > max_count) { if (count > max_count) {
max_count = count; max_count = count;
@ -201,14 +202,16 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), int(max_count / chunk_size), name.c_str()); printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), int(max_count / chunk_size), name.c_str());
} }
} }
gguf_free(ctx_gguf);
ggml_free(ctx);
int m_last_chunk = gguf_get_val_u32(ctx_gguf, chunk_count_idx); int m_last_chunk = gguf_get_val_u32(ctx_gguf, chunk_count_idx);
imatrix_dataset = gguf_get_val_str(ctx_gguf, dataset_idx); imatrix_dataset = gguf_get_val_str(ctx_gguf, dataset_idx);
printf("%s: imatrix dataset='%s'\n", __func__, imatrix_dataset.c_str()); printf("%s: imatrix dataset='%s'\n", __func__, imatrix_dataset.c_str());
printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_chunk); printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_chunk);
gguf_free(ctx_gguf);
ggml_free(ctx);
return m_last_chunk; return m_last_chunk;
} }