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

fix: Only flatten to Q8_0 if the raw target type is a quantization

Browse source 
Branch: GraniteEmbedQuant

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart 2025-01-17 16:34:21 -07:00
parent 614c6e6544
commit 0d7245aa46
1 changed files with 4 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

6
src/llama-quant.cpp
View file

@ -148,6 +148,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
// for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
// with the quantization of the output tensor // with the quantization of the output tensor
const bool is_quantized = ggml_is_quantized(new_type);
if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) { if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
if (qs.params->output_tensor_type < GGML_TYPE_COUNT) { if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
new_type = qs.params->output_tensor_type; new_type = qs.params->output_tensor_type;
@ -155,7 +156,8 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
const int64_t nx = tensor->ne[0]; const int64_t nx = tensor->ne[0];
const int64_t qk_k = ggml_blck_size(new_type); const int64_t qk_k = ggml_blck_size(new_type);
if (arch == LLM_ARCH_FALCON || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE || nx % qk_k != 0) { if (arch == LLM_ARCH_FALCON || nx % qk_k != 0 ||
(is_quantized && (arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE))) {
new_type = GGML_TYPE_Q8_0; new_type = GGML_TYPE_Q8_0;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@ -171,7 +173,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
if (qs.params->token_embedding_type < GGML_TYPE_COUNT) { if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
new_type = qs.params->token_embedding_type; new_type = qs.params->token_embedding_type;
} else { } else {
if (arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) { if (is_quantized && (arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE)) {
new_type = GGML_TYPE_Q8_0; new_type = GGML_TYPE_Q8_0;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||