Compare commits

...

23 Commits

Author SHA1 Message Date
piDack
e76326e3e0
Merge 24bad77ebf into 32d6ee6385 2024-12-24 10:55:02 +08:00
Diego Devesa
32d6ee6385
ggml : fix const usage in SSE path (#10962) 2024-12-23 20:25:52 +01:00
Xuan Son Nguyen
14b699ecde
server : fix missing model id in /model endpoint (#10957)
Some checks are pending
flake8 Lint / Lint (push) Waiting to run
Python Type-Check / pyright type-check (push) Waiting to run
* server : fix missing model id in /model endpoint

* fix ci
2024-12-23 12:52:25 +01:00
Xuan Son Nguyen
485dc01214
server : add system_fingerprint to chat/completion (#10917)
* server : add system_fingerprint to chat/completion

* update README
2024-12-23 12:02:44 +01:00
Radoslav Gerganov
86bf31cfe6
rpc-server : add support for the SYCL backend (#10934) 2024-12-23 10:39:30 +02:00
pcdack
24bad77ebf fix confict 2024-12-19 09:02:35 +00:00
piDack
f91cf62b89
Merge branch 'master' into support_glm_edge_model 2024-12-19 16:50:23 +08:00
piDack
bc93d2a44e
Merge branch 'ggerganov:master' into support_glm_edge_model 2024-12-04 11:50:29 +08:00
liyuhang
6e9fdb0b52 fix chat template 2024-12-04 11:27:01 +08:00
liyuhang
3b409c1e92 add edge template 2024-12-03 19:23:11 +08:00
piDack
82cbfda7b9
Merge branch 'master' into support_glm_edge_model 2024-12-03 13:27:23 +08:00
piDack
5ff563257c
Update src/llama.cpp for not contain <|end|> or </s>
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
2024-11-30 10:29:58 +08:00
piDack
692880535a remove unused AutoTokenizer 2024-11-29 08:15:15 +00:00
piDack
816d93db75 Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model 2024-11-29 08:11:31 +00:00
piDack
6c50e9caca llava.cpp trailing whitespace 2024-11-29 06:11:12 +00:00
piDack
7d80a4aa97 fix format 2024-11-29 06:05:05 +00:00
piDack
55a6f951ca remove debug info 2024-11-29 05:49:13 +00:00
piDack
3b27041727 vision model support 2024-11-26 12:36:48 +00:00
piDack
ae41d3efed Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model 2024-11-26 10:01:09 +00:00
liyuhang
6fc90cb727 support for glm edge model 2024-11-26 09:17:30 +00:00
liyuhang
4f696624a4 use config partial_rotary_factor as rope ratio 2024-11-26 05:29:42 +00:00
liyuhang
a249dc0fbb Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model 2024-11-08 03:53:05 +00:00
liyuhang
677058f470 add glm edge chat model 2024-11-08 03:33:43 +00:00
17 changed files with 640 additions and 85 deletions

View File

@ -91,7 +91,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
@ -111,6 +111,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
- [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
</details>

View File

@ -642,7 +642,7 @@ class Model:
if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
# ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
res = "jina-v2-code"
if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b" or chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
# ref: https://huggingface.co/THUDM/glm-4-9b-chat
res = "chatglm-bpe"
if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
@ -4280,7 +4280,7 @@ class JaisModel(Model):
self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
@Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration")
@Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
class ChatGLMModel(Model):
model_arch = gguf.MODEL_ARCH.CHATGLM
@ -4386,47 +4386,56 @@ class ChatGLMModel(Model):
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
vocab_size = hparams["padded_vocab_size"]
vocab_size = hparams.get("padded_vocab_size",hparams["vocab_size"])
assert max(tokenizer.get_vocab().values()) < vocab_size
tokpre = self.get_vocab_base_pre(tokenizer)
if(hparams["partial_rotary_factor"] == 1.0):
# only for glm-edge series
tokens, toktypes, tokpre = self.get_vocab_base()
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_tokenizer_pre(tokpre)
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
else:
# for glm4 series
tokpre = self.get_vocab_base_pre(tokenizer)
merges = []
vocab = {}
mergeable_ranks = tokenizer._mergeable_ranks
for token, rank in mergeable_ranks.items():
vocab[ChatGLMModel.token_bytes_to_string(token)] = rank
if len(token) == 1:
continue
merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank)
assert len(merged) >= 2 and len(merged) <= 7
merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged)))
merges = []
vocab = {}
mergeable_ranks = tokenizer.mergeable_ranks
for token, rank in mergeable_ranks.items():
vocab[ChatGLMModel.token_bytes_to_string(token)] = rank
if len(token) == 1:
continue
merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank)
assert len(merged) >= 2 and len(merged) <= 7
merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged)))
# for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
added_vocab = tokenizer.get_added_vocab()
reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
# for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
added_vocab = tokenizer.get_added_vocab()
reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.UNUSED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
if tokenizer.added_tokens_decoder[i].special:
toktypes.append(gguf.TokenType.CONTROL)
for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.UNUSED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
if tokenizer.added_tokens_decoder[i].special:
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_tokenizer_pre(tokpre)
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_tokenizer_pre(tokpre)
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
special_vocab.merges = merges
special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
special_vocab.merges = merges
# only add special tokens when they were not already loaded from config.json
special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
@ -4437,16 +4446,20 @@ class ChatGLMModel(Model):
def set_gguf_parameters(self):
n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
n_head_kv = self.hparams.get("multi_query_group_num", n_head)
n_head_kv = self.hparams.get("multi_query_group_num", self.hparams.get("num_key_value_heads", n_head))
self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
self.gguf_writer.add_embedding_length(n_embed)
self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed))
self.gguf_writer.add_block_count(self.hparams["num_layers"])
self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", self.hparams.get("intermediate_size", 4 * n_embed)))
self.gguf_writer.add_block_count(self.hparams.get("num_layers", self.hparams["num_hidden_layers"]))
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head_kv)
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon",1e-5))
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_rope_dimension_count(64)
if "attention_dim" in self.hparams:
rope_dim = self.hparams["attention_dim"]
else:
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
self.gguf_writer.add_add_bos_token(False)
rope_freq = 10000
if "rope_ratio" in self.hparams:
@ -4456,7 +4469,7 @@ class ChatGLMModel(Model):
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
if name.endswith(".rotary_pos_emb.inv_freq"):
if name.endswith(".rotary_pos_emb.inv_freq") or name.startswith("model.vision."):
return []
name = name.removeprefix("transformer.")

View File

@ -0,0 +1,43 @@
# GLMV-EDGE
Currently this implementation supports [glm-edge-v-2b](https://huggingface.co/THUDM/glm-edge-v-2b) and [glm-edge-v-5b](https://huggingface.co/THUDM/glm-edge-v-5b).
## Usage
Build with cmake or run `make llama-llava-cli` to build it.
After building, run: `./llama-llava-cli` to see the usage. For example:
```sh
./llama-llava-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf --image img_path/image.jpg -p "<|system|>\n system prompt <image><|user|>\n prompt <|assistant|>\n"
```
**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
## GGUF conversion
1. Clone a GLMV-EDGE model ([2B](https://huggingface.co/THUDM/glm-edge-v-2b) or [5B](https://huggingface.co/THUDM/glm-edge-v-5b)). For example:
```sh
git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/THUDM/glm-edge-v-2b
```
2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents:
```sh
python ./examples/llava/glmedge-surgery.py -m ../model_path
```
4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF:
```sh
python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
```
5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF:
```sh
python convert_hf_to_gguf.py ../model_path
```
Now both the LLM part and the image encoder are in the `model_path` directory.

View File

@ -101,6 +101,7 @@ static std::string format(const char * fmt, ...) {
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector"
#define KEY_HAS_GLM_PROJ "clip.has_glm_projector"
#define KEY_MINICPMV_VERSION "clip.minicpmv_version"
#define KEY_HAS_QWEN2VL_MERGER "clip.has_qwen2vl_merger"
#define KEY_USE_GELU "clip.use_gelu"
@ -159,6 +160,15 @@ static std::string format(const char * fmt, ...) {
#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
#define TN_MINICPMV_LN "resampler.ln_%s.%s"
#define TN_GLM_ADAPER_CONV "adapter.conv.%s"
#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s"
#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s"
#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
#define TN_GLM_BOI_W "adapter.boi"
#define TN_GLM_EOI_W "adapter.eoi"
enum projector_type {
PROJECTOR_TYPE_MLP,
@ -166,6 +176,7 @@ enum projector_type {
PROJECTOR_TYPE_LDP,
PROJECTOR_TYPE_LDPV2,
PROJECTOR_TYPE_RESAMPLER,
PROJECTOR_TYPE_ADAPTER,
PROJECTOR_TYPE_MERGER,
PROJECTOR_TYPE_UNKNOWN,
};
@ -175,6 +186,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_LDPV2, "ldpv2"},
{ PROJECTOR_TYPE_RESAMPLER, "resampler"},
{ PROJECTOR_TYPE_ADAPTER, "adapter"},
{ PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
};
@ -499,6 +511,12 @@ struct clip_vision_model {
struct ggml_tensor * mm_4_w = NULL;
struct ggml_tensor * mm_4_b = NULL;
//GLMV-Edge projection
struct ggml_tensor * mm_model_adapter_conv_w;
struct ggml_tensor * mm_model_adapter_conv_b;
struct ggml_tensor * boi_w;
struct ggml_tensor * eoi_w;
// MobileVLM projection
struct ggml_tensor * mm_model_mlp_1_w;
struct ggml_tensor * mm_model_mlp_1_b;
@ -559,6 +577,7 @@ struct clip_ctx {
bool has_vision_encoder = false;
bool has_llava_projector = false;
bool has_minicpmv_projector = false;
bool has_glm_projector = false;
bool has_qwen2vl_merger = false;
int minicpmv_version = 2;
@ -637,7 +656,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
const int batch_size = imgs->size;
if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) {
GGML_ASSERT(batch_size == 1);
}
@ -730,8 +749,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
}
// loop over layers
if (ctx->has_minicpmv_projector || ctx->has_qwen2vl_merger) {
// TODO: figure out why we doing thing in this way ???
if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
n_layer += 1;
}
for (int il = 0; il < n_layer - 1; il++) {
@ -1086,7 +1104,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
GGML_ASSERT(false);
}
}
else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
// glm projector
else if(ctx->has_glm_projector){
if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
embeddings = ggml_reshape_3d(ctx0,embeddings,gridsz,gridsz,embeddings->ne[1]);
embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
//GLU
{
embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
embeddings = ggml_norm(ctx0, embeddings, eps);
embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
embeddings = ggml_gelu_inplace(ctx0, embeddings);
struct ggml_tensor * x = embeddings;
embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
x = ggml_mul_mat(ctx0,model.mm_model_mlp_1_w,x);
embeddings = ggml_silu_inplace(ctx0,embeddings);
embeddings = ggml_mul(ctx0,embeddings,x);
embeddings = ggml_mul_mat(ctx0,model.mm_model_mlp_3_w,embeddings);
}
}else{
GGML_ABORT("fatel error");
}
}else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
@ -1275,6 +1319,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
}
idx = gguf_find_key(ctx, KEY_HAS_GLM_PROJ);
if (idx != -1) {
new_clip->has_glm_projector = gguf_get_val_bool(ctx, idx);
}
idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER);
if (idx != -1) {
new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx);
@ -1299,6 +1348,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
LOG_INF("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
LOG_INF("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector);
LOG_INF("%s: glm_projector: %d\n", __func__, new_clip->has_glm_projector);
LOG_INF("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
LOG_INF("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
}
@ -1566,6 +1616,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
}
else if(new_clip->proj_type == PROJECTOR_TYPE_ADAPTER){
vision_model.mm_model_adapter_conv_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "weight"));
vision_model.mm_model_adapter_conv_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "bias"));
vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_LINEAR,"weight"));
vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"weight"));
vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"bias"));
vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_H_2_4H,"weight"));
vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_GATE,"weight"));
vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_4H_2_H,"weight"));
vision_model.boi_w = get_tensor(new_clip->ctx_data, TN_GLM_BOI_W);
vision_model.eoi_w = get_tensor(new_clip->ctx_data, TN_GLM_EOI_W);
}
else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) {
vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
@ -2098,6 +2160,20 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
return true;
}
if(ctx->has_glm_projector){
res_imgs->size = 1;
res_imgs->data = new clip_image_f32[res_imgs->size];
clip_image_u8 resized_image;
int32_t sz=ctx->vision_model.hparams.image_size;
bicubic_resize(*img, resized_image,sz,sz);
clip_image_f32 * res = clip_image_f32_init();
//clip_image_save_to_bmp(resized_image, "resized.bmp");
normalize_image_u8_to_f32(&resized_image, res, ctx->image_mean, ctx->image_std);
res_imgs->data[0] = *res;
clip_image_f32_free(res);
return true;
}
bool pad_to_square = true;
if (!ctx->has_vision_encoder) {
LOG_ERR("This gguf file seems to have no vision encoder\n");
@ -2283,6 +2359,8 @@ void clip_free(clip_ctx * ctx) {
}
size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
if(ctx->has_glm_projector)
return (clip_n_patches(ctx)+2) * clip_n_mmproj_embd(ctx) * sizeof(float);
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
}
@ -2325,7 +2403,7 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_ADAPTER) {
n_patches /= 4;
} else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
if (ctx->minicpmv_version == 2) {
@ -2455,6 +2533,12 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
if (ctx->has_minicpmv_projector) {
GGML_ASSERT(batch_size == 1);
}
if(ctx->has_glm_projector){
GGML_ASSERT(batch_size == 1);
ggml_tensor * boi = ctx->vision_model.boi_w;
ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi));
vec=(float*)(vec+ggml_nelements(boi)); //offset for boi
}
// build the inference graph
ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
@ -2604,7 +2688,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
free(positions_data);
{
if (!ctx->has_glm_projector){
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
int* patches_data = (int*)malloc(ggml_nbytes(patches));
for (int i = 0; i < num_patches; i++) {
@ -2628,6 +2712,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// copy the embeddings to the location passed by the user
ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
if(ctx->has_glm_projector){
//eoi
ggml_tensor * eoi = ctx->vision_model.eoi_w;
int offset=ggml_nelements(eoi)*clip_n_patches(ctx);
ggml_backend_tensor_get(eoi,vec+offset,0,ggml_nbytes(eoi));
}
return true;
}
@ -2785,6 +2876,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return 3584;
}
}
if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
return ctx->vision_model.mm_model_mlp_3_w->ne[1];
}
if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
return ctx->vision_model.mm_1_b->ne[0];
}
@ -2800,6 +2894,9 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
return 0;
}
bool clip_is_glm(const struct clip_ctx * ctx) {
return ctx->has_glm_projector;
}
bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
return ctx->has_qwen2vl_merger;
}

View File

@ -93,6 +93,8 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
#ifdef __cplusplus
}
#endif

View File

@ -0,0 +1,280 @@
import argparse
import os
import json
import re
import torch
import numpy as np
from gguf import *
TEXT = "clip.text"
VISION = "clip.vision"
from transformers import SiglipVisionModel, SiglipVisionConfig
def k(raw_key: str, arch: str) -> str:
return raw_key.format(arch=arch)
def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
if name in (
"logit_scale",
"text_model.embeddings.position_ids",
"vision_model.embeddings.position_ids",
):
return True
if name in (
"vision_model.head.probe",
"vision_model.head.attention.in_proj_weight",
"vision_model.head.attention.in_proj_bias",
"vision_model.head.attention.out_proj.weight",
"vision_model.head.attention.out_proj.bias",
"vision_model.head.layernorm.weight",
"vision_model.head.layernorm.bias",
"vision_model.head.mlp.fc1.weight",
"vision_model.head.mlp.fc1.bias",
"vision_model.head.mlp.fc2.weight",
"vision_model.head.mlp.fc2.bias"
):
return True
if name.startswith("v") and not has_vision:
return True
if name.startswith("t") and not has_text:
return True
return False
def get_tensor_name(name: str) -> str:
if "projection" in name:
return name
if "mm_projector" in name:
name = name.replace("model.mm_projector", "mm")
name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
return name
return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = (
list(range(ord("!"), ord("~") + 1))
+ list(range(ord("¡"), ord("¬") + 1))
+ list(range(ord("®"), ord("ÿ") + 1))
)
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8 + n)
n += 1
cs = [chr(n) for n in cs]
return dict(zip(bs, cs))
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
ap.add_argument("--text-only", action="store_true", required=False,
help="Save a text-only model. It can't be used to encode images")
ap.add_argument("--vision-only", action="store_true", required=False,
help="Save a vision-only model. It can't be used to encode texts")
ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
help="The clip model is from openclip (for ViT-SO400M type))")
ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2","adapter"], default="adapter")
ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
default_image_mean = [0.5, 0.5, 0.5]
default_image_std = [0.5, 0.5, 0.5]
ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
# with proper
args = ap.parse_args()
if args.text_only and args.vision_only:
print("--text-only and --image-only arguments cannot be specified at the same time.")
exit(1)
if args.use_f32:
print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
# output in the same directory as the model if output_dir is None
dir_model = args.model_dir
if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
vocab = None
tokens = None
else:
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
vocab = json.load(f)
tokens = [key for key in vocab]
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
config = json.load(f)
if args.clip_model_is_vision:
v_hparams = config
t_hparams = None
else:
v_hparams = config["vision_config"]
t_hparams = None
# possible data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if args.use_f32:
ftype = 0
vision_config = SiglipVisionConfig(**v_hparams)
model = SiglipVisionModel(vision_config)
model.load_state_dict(torch.load(os.path.join(dir_model, "glm.clip")))
fname_middle = None
has_text_encoder = False
has_vision_encoder = True
has_glm_projector = True
if args.text_only:
fname_middle = "text-"
has_vision_encoder = False
elif args.llava_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_glm_projector = True
elif args.vision_only:
fname_middle = "vision-"
has_text_encoder = False
else:
fname_middle = ""
output_dir = args.output_dir if args.output_dir is not None else dir_model
os.makedirs(output_dir, exist_ok=True)
output_prefix = os.path.basename(output_dir).replace("ggml_", "")
fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
fout = GGUFWriter(path=fname_out, arch="clip")
fout.add_bool("clip.has_text_encoder", has_text_encoder)
fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
fout.add_bool("clip.has_glm_projector", has_glm_projector)
fout.add_file_type(ftype)
model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
fout.add_name(model_name)
if has_glm_projector:
fout.add_description("image encoder for glm4v")
fout.add_string("clip.projector_type", "adapter")
else:
fout.add_description("two-tower CLIP model")
if has_text_encoder:
assert t_hparams is not None
assert tokens is not None
# text_model hparams
fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
fout.add_token_list(tokens)
if has_vision_encoder:
# vision_model hparams
fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
fout.add_uint32("clip.vision.projection_dim", 0)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), v_hparams["num_hidden_layers"])
image_mean = args.image_mean if args.image_mean is not None else default_image_mean
image_std = args.image_std if args.image_std is not None else default_image_std
fout.add_array("clip.vision.image_mean", image_mean)
fout.add_array("clip.vision.image_std", image_std)
fout.add_bool("clip.use_gelu", True)
if has_glm_projector:
# model.vision_model.encoder.layers.pop(-1) # pyright: ignore[reportAttributeAccessIssue]
projector = torch.load(args.llava_projector)
for name, data in projector.items():
name = get_tensor_name(name)
# pw and dw conv ndim==4
if data.ndim == 2 or data.ndim == 4:
data = data.squeeze().numpy().astype(np.float16)
else:
data = data.squeeze().numpy().astype(np.float32)
if name.startswith("vision."):
name=name.replace("vision.","")
fout.add_tensor(name, data)
print(f"Projector {name} - {data.dtype} - shape = {data.shape}")
# print(f"Projector {name} tensors added\n")
state_dict = model.state_dict() # pyright: ignore[reportAttributeAccessIssue]
for name, data in state_dict.items():
if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_glm_projector):
# we don't need this
print(f"skipping parameter: {name}")
continue
name = get_tensor_name(name)
data = data.squeeze().numpy()
n_dims = len(data.shape)
# ftype == 0 -> float32, ftype == 1 -> float16
ftype_cur = 0
if n_dims == 4:
print(f"tensor {name} is always saved in f16")
data = data.astype(np.float16)
ftype_cur = 1
elif ftype == 1:
if name[-7:] == ".weight" and n_dims == 2:
# print(" Converting to float16")
data = data.astype(np.float16)
ftype_cur = 1
else:
# print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
else:
if data.dtype != np.float32:
# print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
print(f"siglip {name} - {data.dtype} - shape = {data.shape}")
# print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
fout.add_tensor(name, data)
fout.write_header_to_file()
fout.write_kv_data_to_file()
fout.write_tensors_to_file()
fout.close()
print("Done. Output file: " + fname_out)

View File

@ -0,0 +1,33 @@
import argparse
import os
import torch
from transformers import AutoModel
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", help="Path to GLM model")
args = ap.parse_args()
# find the model part that includes the the multimodal projector weights
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
checkpoint = model.state_dict()
# get a list of mm tensor names
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.adapter.")]
# store these tensors in a new dictionary and torch.save them
projector = {name: checkpoint[name].float() for name in mm_tensors}
torch.save(projector, f"{args.model}/glm.projector")
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.vit.model.vision_model.")]
if len(clip_tensors) > 0:
clip = {name.replace("vision.vit.model.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/glm.clip")
# added tokens should be removed to be able to convert Mistral models
if os.path.exists(f"{args.model}/added_tokens.json"):
with open(f"{args.model}/added_tokens.json", "w") as f:
f.write("{}\n")
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
print(f"Also, use {args.model}glm.projector to prepare a glm-encoder.gguf file.")

View File

@ -314,6 +314,20 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
clip_add_load_image_size(ctx_clip, load_image_size);
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
}
else if (clip_is_glm(ctx_clip)){
struct clip_image_size * load_image_size = clip_image_size_init();
load_image_size->width = img_res_v.data[0].nx;
load_image_size->height = img_res_v.data[0].ny;
clip_add_load_image_size(ctx_clip, load_image_size);
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
*n_img_pos = (pos * pos + 2);
if (!encoded){
LOG_ERR("Unable to encode image \n");
return false;
}
}
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
*n_img_pos = clip_n_patches(ctx_clip);
@ -398,6 +412,9 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
if (clip_is_minicpmv(ctx_clip)) {
num_max_patches = 10;
}
if (clip_is_glm(ctx_clip)) {
num_max_patches = 1;
}
float * image_embd;
if (clip_is_qwen2vl(ctx_clip)) {
// qwen2vl don't split image into chunks, so `num_max_patches` is not needed.

View File

@ -12,6 +12,10 @@
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#include "ggml-rpc.h"
#ifdef _WIN32
# include <windows.h>
@ -91,6 +95,12 @@ static ggml_backend_t create_backend() {
if (!backend) {
fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
}
#elif GGML_USE_SYCL
fprintf(stderr, "%s: using SYCL backend\n", __func__);
backend = ggml_backend_sycl_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
}
#endif
// if there aren't GPU Backends fallback to CPU backend
@ -106,6 +116,8 @@ static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
#elif GGML_USE_VULKAN
ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
#elif GGML_USE_SYCL
ggml_backend_sycl_get_device_memory(0, free_mem, total_mem);
#else
#ifdef _WIN32
MEMORYSTATUSEX status;

View File

@ -724,7 +724,8 @@ This endpoint is public (no API key check). By default, it is read-only. To make
},
"total_slots": 1,
"model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
"chat_template": "..."
"chat_template": "...",
"build_info": "b(build number)-(build commit hash)"
}
```

View File

@ -595,10 +595,11 @@ struct server_task_result_cmpl_final : server_task_result {
std::time_t t = std::time(0);
json res = json {
{"choices", json::array({choice})},
{"created", t},
{"model", oaicompat_model},
{"object", "chat.completion"},
{"choices", json::array({choice})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
@ -632,11 +633,12 @@ struct server_task_result_cmpl_final : server_task_result {
};
json ret = json {
{"choices", json::array({choice})},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"object", "chat.completion.chunk"},
{"choices", json::array({choice})},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion.chunk"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
@ -761,11 +763,12 @@ struct server_task_result_cmpl_partial : server_task_result {
}
json ret = json {
{"choices", choices},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"object", "chat.completion.chunk"}
{"choices", choices},
{"created", t},
{"id", oaicompat_cmpl_id},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "chat.completion.chunk"}
};
if (timings.prompt_n >= 0) {
@ -3476,6 +3479,7 @@ int main(int argc, char ** argv) {
{ "total_slots", ctx_server.params_base.n_parallel },
{ "model_path", ctx_server.params_base.model },
{ "chat_template", llama_get_chat_template(ctx_server.model) },
{ "build_info", build_info },
};
res_ok(res, data);
@ -3697,7 +3701,7 @@ int main(int argc, char ** argv) {
{"object", "list"},
{"data", {
{
{"id", params.model_alias},
{"id", params.model_alias.empty() ? params.model : params.model_alias},
{"object", "model"},
{"created", std::time(0)},
{"owned_by", "llamacpp"},

View File

@ -31,6 +31,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
})
assert res.status_code == 200
assert "cmpl" in res.body["id"] # make sure the completion id has the expected format
assert res.body["system_fingerprint"].startswith("b")
assert res.body["model"] == model if model is not None else server.model_alias
assert res.body["usage"]["prompt_tokens"] == n_prompt
assert res.body["usage"]["completion_tokens"] == n_predicted
@ -63,6 +64,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
last_cmpl_id = None
for data in res:
choice = data["choices"][0]
assert data["system_fingerprint"].startswith("b")
assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
if last_cmpl_id is None:
last_cmpl_id = data["id"]
@ -92,6 +94,7 @@ def test_chat_completion_with_openai_library():
seed=42,
temperature=0.8,
)
assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
assert res.choices[0].finish_reason == "length"
assert res.choices[0].message.content is not None
assert match_regex("(Suddenly)+", res.choices[0].message.content)

View File

@ -56,6 +56,8 @@ static T json_value(const json & body, const std::string & key, const T & defaul
}
}
const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
//
// tokenizer and input processing utils
//

View File

@ -986,7 +986,7 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
#define GGML_F16_STEP 32
#define GGML_F16_EPR 4
static inline __m128 __sse_f16x4_load(ggml_fp16_t *x) {
static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]);
@ -997,7 +997,7 @@ static inline __m128 __sse_f16x4_load(ggml_fp16_t *x) {
return _mm_loadu_ps(tmp);
}
static inline void __sse_f16x4_store(ggml_fp16_t *x, __m128 y) {
static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
float arr[4];
_mm_storeu_ps(arr, y);

View File

@ -1284,6 +1284,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,

View File

@ -1440,6 +1440,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
@ -1715,6 +1718,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_LLAMA_3,
LLM_CHAT_TEMPLATE_CHATGML_3,
LLM_CHAT_TEMPLATE_CHATGML_4,
LLM_CHAT_TEMPLATE_GLMEDGE,
LLM_CHAT_TEMPLATE_MINICPM,
LLM_CHAT_TEMPLATE_EXAONE_3,
LLM_CHAT_TEMPLATE_RWKV_WORLD,
@ -1749,6 +1753,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 },
{ "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 },
{ "chatglm4", LLM_CHAT_TEMPLATE_CHATGML_4 },
{ "glmedge", LLM_CHAT_TEMPLATE_GLMEDGE },
{ "minicpm", LLM_CHAT_TEMPLATE_MINICPM },
{ "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 },
{ "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD },
@ -6300,8 +6305,20 @@ static void llm_load_hparams(
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 28: model.type = e_model::MODEL_6B; break;
case 40: model.type = e_model::MODEL_9B; break;
case 28: {
if(hparams.n_head(0)==16){
model.type = e_model::MODEL_1_6B;
}else{
model.type = e_model::MODEL_6B;
}
}break;
case 40:{
if(hparams.n_head(0)==24){
model.type = e_model::MODEL_4B;
}else{
model.type = e_model::MODEL_9B;
}
} break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
@ -9434,9 +9451,14 @@ static bool llm_load_tensors(
auto & layer = model.layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
if(model.type == e_model::MODEL_1_6B || model.type == e_model::MODEL_4B){
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
}else{
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
}
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
@ -16827,20 +16849,28 @@ struct llm_build_context {
struct ggml_tensor * Qcur = nullptr;
struct ggml_tensor * Kcur = nullptr;
struct ggml_tensor * Vcur = nullptr;
if(model.type == e_model::MODEL_1_6B || model.type == e_model::MODEL_4B){
Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
}else{
cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
if(model.layers[il].bqkv){
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
}
Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
}
cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
//printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
@ -22921,6 +22951,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_CHATGML_3;
} else if (tmpl_contains("[gMASK]<sop>")) {
return LLM_CHAT_TEMPLATE_CHATGML_4;
} else if (tmpl_contains("<|user|>") && tmpl_contains("<|assistant|>") && !tmpl_contains("<|end|>") && !tmpl_contains("</s>")) {
return LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains(LU8("<用户>"))) {
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
return LLM_CHAT_TEMPLATE_MINICPM;
@ -23204,6 +23236,14 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<|assistant|>";
}
} else if(tmpl == LLM_CHAT_TEMPLATE_GLMEDGE){
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>" << "\n" << message->content;
}
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
for (auto message : chat) {

View File

@ -61,6 +61,8 @@ int main(void) {
"{% for message in messages %}{% if loop.first %}[gMASK]sop<|{{ message['role'] }}|>\n {{ message['content'] }}{% else %}<|{{ message['role'] }}|>\n {{ message['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
// ChatGLM4
u8"[gMASK]<sop>{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的,你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
// GLM-edge
"{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>\n",
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
u8"{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
// DeepSeek-V2
@ -119,6 +121,8 @@ int main(void) {
"[gMASK]sop<|system|>\n You are a helpful assistant<|user|>\n Hello<|assistant|>\n Hi there<|user|>\n Who are you<|assistant|>\n I am an assistant <|user|>\n Another question<|assistant|>",
// ChatGLM4
"[gMASK]<sop><|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
// GLM-Edge
"<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
u8"You are a helpful assistant<用户>Hello<AI>Hi there<用户>Who are you<AI>I am an assistant<用户>Another question<AI>",
// DeepSeek-V2