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

quantize: add imatrix and dataset metadata in GGUF (#6658)

Browse source 
* imatrix: save the dataset file used in the output file

* llama: support kv overrides type string string

* common: factorize KV Overrides parsing between common and server

* quantize: add imatrix n entries and dataset KV metadata
quantize: factorize KV Overrides parsing between common
#6656

* llama: remove kv override str_value initialization as it does not compile on some toolchain

* quantize: add imatrix m_last_call as `quantize.imatrix.chunks_count`

* quantize: add imatrix filename in KV

* llama: add llama_model_kv_override_free

* common: add llama_model_kv_override_free
common: free kv override if used after model loading

* llama: finally move the string KV override value to the stack

* llama : minor

* no need to add a NUL to the std::vector, std::string can be initialized from a pair of iterators.

Co-authored-by: slaren <slarengh@gmail.com>

* kv override: ensure string termination

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Pierrick Hymbert 2024-04-26 20:06:33 +02:00 • committed by GitHub
parent 017e6999b5
commit 0c4d489e29
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 186 additions and 171 deletions
Download patch file Download diff file Expand all files Collapse all files

35
llama.cpp
View file

@ -2883,6 +2883,7 @@ namespace GGUFMeta {
case LLAMA_KV_OVERRIDE_TYPE_BOOL: return "bool";
case LLAMA_KV_OVERRIDE_TYPE_INT: return "int";
case LLAMA_KV_OVERRIDE_TYPE_FLOAT: return "float";
case LLAMA_KV_OVERRIDE_TYPE_STR: return "str";
}
return "unknown";
}
@ -2894,13 +2895,16 @@ namespace GGUFMeta {
__func__, override_type_to_str(ovrd->tag), ovrd->key);
switch (ovrd->tag) {
case LLAMA_KV_OVERRIDE_TYPE_BOOL: {
LLAMA_LOG_INFO("%s\n", ovrd->bool_value ? "true" : "false");
LLAMA_LOG_INFO("%s\n", ovrd->val_bool ? "true" : "false");
} break;
case LLAMA_KV_OVERRIDE_TYPE_INT: {
LLAMA_LOG_INFO("%" PRId64 "\n", ovrd->int_value);
LLAMA_LOG_INFO("%" PRId64 "\n", ovrd->val_i64);
} break;
case LLAMA_KV_OVERRIDE_TYPE_FLOAT: {
LLAMA_LOG_INFO("%.6f\n", ovrd->float_value);
LLAMA_LOG_INFO("%.6f\n", ovrd->val_f64);
} break;
case LLAMA_KV_OVERRIDE_TYPE_STR: {
LLAMA_LOG_INFO("%s\n", ovrd->val_str);
} break;
default:
// Shouldn't be possible to end up here, but just in case...
@ -2919,7 +2923,7 @@ namespace GGUFMeta {
static typename std::enable_if<std::is_same<OT, bool>::value, bool>::type
try_override(OT & target, const struct llama_model_kv_override * ovrd) {
if (validate_override(LLAMA_KV_OVERRIDE_TYPE_BOOL, ovrd)) {
target = ovrd->bool_value;
target = ovrd->val_bool;
return true;
}
return false;
@ -2929,7 +2933,7 @@ namespace GGUFMeta {
static typename std::enable_if<!std::is_same<OT, bool>::value && std::is_integral<OT>::value, bool>::type
try_override(OT & target, const struct llama_model_kv_override * ovrd) {
if (validate_override(LLAMA_KV_OVERRIDE_TYPE_INT, ovrd)) {
target = ovrd->int_value;
target = ovrd->val_i64;
return true;
}
return false;
@ -2939,7 +2943,7 @@ namespace GGUFMeta {
static typename std::enable_if<std::is_floating_point<OT>::value, bool>::type
try_override(T & target, const struct llama_model_kv_override * ovrd) {
if (validate_override(LLAMA_KV_OVERRIDE_TYPE_FLOAT, ovrd)) {
target = ovrd->float_value;
target = ovrd->val_f64;
return true;
}
return false;
@ -2948,12 +2952,11 @@ namespace GGUFMeta {
template<typename OT>
static typename std::enable_if<std::is_same<OT, std::string>::value, bool>::type
try_override(T & target, const struct llama_model_kv_override * ovrd) {
(void)target;
(void)ovrd;
if (!ovrd) { return false; }
// Currently, we should never end up here so it would be a bug if we do.
throw std::runtime_error(format("Unsupported attempt to override string type for metadata key %s\n",
ovrd ? ovrd->key : "NULL"));
if (validate_override(LLAMA_KV_OVERRIDE_TYPE_STR, ovrd)) {
target = ovrd->val_str;
return true;
}
return false;
}
static bool set(const gguf_context * ctx, const int k, T & target, const struct llama_model_kv_override * ovrd = nullptr) {
@ -14548,11 +14551,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
for (auto & o : overrides) {
if (o.key[0] == 0) break;
if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
gguf_set_val_f32(ctx_out, o.key, o.float_value);
gguf_set_val_f32(ctx_out, o.key, o.val_f64);
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
gguf_set_val_i32(ctx_out, o.key, o.int_value);
gguf_set_val_i32(ctx_out, o.key, o.val_i64);
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
gguf_set_val_bool(ctx_out, o.key, o.bool_value);
gguf_set_val_bool(ctx_out, o.key, o.val_bool);
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
gguf_set_val_str(ctx_out, o.key, o.val_str);
} else {
LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key);
}