nodes/city96-ComfyUI-GGUF
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'main' into stable

Browse Source
This commit is contained in:
City
2025-08-05 14:30:37 +02:00
parent c51a948233 79ba1d4b17
commit bf846d54a8
4 changed files with 65 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
loader.py
+1 -1
View File
@@ -7,7 +7,7 @@ import gguf
from .ops import GGMLTensor
from .dequant import is_quantized, dequantize_tensor
IMG_ARCH_LIST = {"flux", "sd1", "sdxl", "sd3", "aura", "hidream", "cosmos", "ltxv", "hyvid", "wan"}
IMG_ARCH_LIST = {"flux", "sd1", "sdxl", "sd3", "aura", "hidream", "cosmos", "ltxv", "hyvid", "wan", "lumina2", "qwen_image"}
TXT_ARCH_LIST = {"t5", "t5encoder", "llama"}
def get_orig_shape(reader, tensor_name):
tools/README.md
+27
View File
@@ -40,6 +40,8 @@ git apply ..\lcpp.patch
Compile the llama-quantize binary. This example uses cmake, on linux you can just use make.
### Visual Studio 2019, Linux, etc...
```
mkdir build
cmake -B build
@@ -47,6 +49,31 @@ cmake --build build --config Debug -j10 --target llama-quantize
cd ..
```
### Visual Studio 2022
```
mkdir build
cmake -B build -DCMAKE_CXX_STANDARD=17 -DCMAKE_CXX_STANDARD_REQUIRED=ON -DCMAKE_CXX_FLAGS="-std=c++17"
```
Edit the `llama.cpp\common\log.cpp` file, inserts two lines after the existing first line:
```
#include "log.h"
#define _SILENCE_CXX23_CHRONO_DEPRECATION_WARNING
#include <chrono>
```
Then you can build the project:
```
cmake --build build --config Debug -j10 --target llama-quantize
cd ..
```
### Quantize your model
Now you can use the newly build binary to quantize your model to the desired format:
```
llama.cpp\build\bin\Debug\llama-quantize.exe E:\models\unet\flux1-dev-BF16.gguf E:\models\unet\flux1-dev-Q4_K_S.gguf Q4_K_S
tools/convert.py
+9 -2
View File
@@ -139,8 +139,14 @@ class ModelSD1(ModelTemplate):
), # Non-diffusers
]
# The architectures are checked in order and the first successful match terminates the search.
arch_list = [ModelFlux, ModelSD3, ModelAura, ModelHiDream, CosmosPredict2, ModelLTXV, ModelHyVid, ModelWan, ModelSDXL, ModelSD1]
class ModelLumina2(ModelTemplate):
arch = "lumina2"
keys_detect = [
("cap_embedder.1.weight", "context_refiner.0.attention.qkv.weight")
]
arch_list = [ModelFlux, ModelSD3, ModelAura, ModelHiDream, CosmosPredict2,
ModelLTXV, ModelHyVid, ModelWan, ModelSDXL, ModelSD1, ModelLumina2]
def is_model_arch(model, state_dict):
# check if model is correct
@@ -356,3 +362,4 @@ def convert_file(path, dst_path=None, interact=True, overwrite=False):
if __name__ == "__main__":
args = parse_args()
convert_file(args.src, args.dst)
tools/lcpp.patch
+28 -13
View File
@@ -39,10 +39,10 @@ index b16c462f..6d1568f1 100644
const int idx = gguf_find_tensor(ctx, name);
if (idx < 0) {
diff --git a/src/llama.cpp b/src/llama.cpp
index 24e1f1f0..9957ea30 100644
index 24e1f1f0..25db4c69 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -205,6 +205,16 @@ enum llm_arch {
@@ -205,6 +205,17 @@ enum llm_arch {
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_CHAMELEON,
@@ -56,10 +56,11 @@ index 24e1f1f0..9957ea30 100644
+ LLM_ARCH_WAN,
+ LLM_ARCH_HIDREAM,
+ LLM_ARCH_COSMOS,
+ LLM_ARCH_LUMINA2,
LLM_ARCH_UNKNOWN,
};
@@ -258,6 +268,16 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
@@ -258,6 +269,17 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GRANITE, "granite" },
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
{ LLM_ARCH_CHAMELEON, "chameleon" },
@@ -73,10 +74,11 @@ index 24e1f1f0..9957ea30 100644
+ { LLM_ARCH_WAN, "wan" },
+ { LLM_ARCH_HIDREAM, "hidream" },
+ { LLM_ARCH_COSMOS, "cosmos" },
+ { LLM_ARCH_LUMINA2, "lumina2" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@@ -1531,6 +1551,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
@@ -1531,6 +1553,17 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
},
},
@@ -90,10 +92,11 @@ index 24e1f1f0..9957ea30 100644
+ { LLM_ARCH_WAN, {}},
+ { LLM_ARCH_HIDREAM, {}},
+ { LLM_ARCH_COSMOS, {}},
+ { LLM_ARCH_LUMINA2, {}},
{
LLM_ARCH_UNKNOWN,
{
@@ -5403,6 +5433,24 @@ static void llm_load_hparams(
@@ -5403,6 +5436,25 @@ static void llm_load_hparams(
// get general kv
ml.get_key(LLM_KV_GENERAL_NAME, model.name, false);
@@ -109,6 +112,7 @@ index 24e1f1f0..9957ea30 100644
+ case LLM_ARCH_WAN:
+ case LLM_ARCH_HIDREAM:
+ case LLM_ARCH_COSMOS:
+ case LLM_ARCH_LUMINA2:
+ model.ftype = ml.ftype;
+ return;
+ default:
@@ -118,7 +122,7 @@ index 24e1f1f0..9957ea30 100644
// get hparams kv
ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab);
@@ -18016,6 +18064,132 @@ static void llama_tensor_dequantize_internal(
@@ -18016,6 +18068,134 @@ static void llama_tensor_dequantize_internal(
workers.clear();
}
@@ -174,7 +178,8 @@ index 24e1f1f0..9957ea30 100644
+ ++qs.i_attention_wv;
+ } else if ( // Rules for fused qkv attention
+ (name.find("attn_qkv.weight") != std::string::npos) ||
+ (name.find("attn.qkv.weight") != std::string::npos)
+ (name.find("attn.qkv.weight") != std::string::npos) ||
+ (name.find("attention.qkv.weight") != std::string::npos)
+ ) {
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+ new_type = GGML_TYPE_Q4_K;
@@ -191,7 +196,8 @@ index 24e1f1f0..9957ea30 100644
+ (name.find(".ffn.2.weight") != std::string::npos) || // is this even the right way around?
+ (name.find(".ff.net.2.weight") != std::string::npos) ||
+ (name.find(".mlp.layer2.weight") != std::string::npos) ||
+ (name.find(".adaln_modulation_mlp.2.weight") != std::string::npos)
+ (name.find(".adaln_modulation_mlp.2.weight") != std::string::npos) ||
+ (name.find(".feed_forward.w2.weight") != std::string::npos)
+ ) {
+ // TODO: add back `layer_info` with some model specific logic + logic further down
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
@@ -251,7 +257,7 @@ index 24e1f1f0..9957ea30 100644
static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
const std::string name = ggml_get_name(tensor);
@@ -18513,7 +18687,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
@@ -18513,7 +18693,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (llama_model_has_encoder(&model)) {
n_attn_layer *= 3;
}
@@ -262,7 +268,7 @@ index 24e1f1f0..9957ea30 100644
}
size_t total_size_org = 0;
@@ -18547,6 +18723,51 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
@@ -18547,6 +18729,51 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
ctx_outs[i_split] = gguf_init_empty();
}
gguf_add_tensor(ctx_outs[i_split], tensor);
@@ -314,7 +320,7 @@ index 24e1f1f0..9957ea30 100644
}
// Set split info if needed
@@ -18647,6 +18868,101 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
@@ -18647,6 +18874,110 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos;
@@ -408,6 +414,15 @@ index 24e1f1f0..9957ea30 100644
+ quantize &= name.find("pos_embedder.") == std::string::npos;
+ quantize &= name.find("final_layer.") == std::string::npos;
+ }
+ if (model.arch == LLM_ARCH_LUMINA2) {
+ image_model = true;
+ quantize &= name.find("t_embedder.") == std::string::npos;
+ quantize &= name.find("x_embedder.") == std::string::npos;
+ quantize &= name.find("final_layer.") == std::string::npos;
+ quantize &= name.find("cap_embedder.") == std::string::npos;
+ quantize &= name.find("context_refiner.") == std::string::npos;
+ quantize &= name.find("noise_refiner.") == std::string::npos;
+ }
+ // ignore 3D/4D tensors for image models as the code was never meant to handle these
+ if (image_model) {
+ quantize &= ggml_n_dims(tensor) == 2;
@@ -416,7 +431,7 @@ index 24e1f1f0..9957ea30 100644
enum ggml_type new_type;
void * new_data;
size_t new_size;
@@ -18655,6 +18971,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
@@ -18655,6 +18986,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
new_type = default_type;
// get more optimal quantization type based on the tensor shape, layer, etc.
@@ -426,7 +441,7 @@ index 24e1f1f0..9957ea30 100644
if (!params->pure && ggml_is_quantized(default_type)) {
new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
}
@@ -18664,6 +18983,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
@@ -18664,6 +18998,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) {
new_type = params->output_tensor_type;
}