Skip to content

Add Phi-4-mini-instruct support #12099

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

Add Phi-4-mini-instruct support #12099

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1837951
Added Phi-4-mini-instruct support
nisparks Feb 27, 2025
3968c5a
Update regex per ngxson
nisparks Feb 27, 2025
958c7ca
Change the vocab base to Xenova/gpt-4o
nisparks Feb 27, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 20 additions & 3 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -699,6 +699,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5":
# ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
res = "deepseek-r1-qwen"
if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
# ref: https://huggingface.co/Xenova/gpt-4o
res = "gpt-4o"

if res is None:
logger.warning("\n")
Expand Down Expand Up @@ -2505,14 +2508,18 @@ def set_vocab(self):

def set_gguf_parameters(self):
block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
if self.hparams.get("partial_rotary_factor") is not None:
rot_pct = self.find_hparam(["partial_rotary_factor"])
else:
rot_pct = 1.0

n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
rms_eps = self.find_hparam(["rms_norm_eps"])
max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
rope_dims = n_embd // n_head
rope_dims = int(rot_pct * n_embd) // n_head

self.gguf_writer.add_context_length(max_pos_embds)
self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
Expand All @@ -2530,13 +2537,23 @@ def set_gguf_parameters(self):
if sliding_window is None:
sliding_window = 0
self.gguf_writer.add_sliding_window(sliding_window)
if self.hparams.get("rope_scaling") is not None:
if self.hparams["rope_scaling"].get("type") == "longrope":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")


def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
if self.hparams.get("partial_rotary_factor") is not None:
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This can be written shorter: self.hparams.get("partial_rotary_factor", 1.0)

rot_pct = self.find_hparam(["partial_rotary_factor"])
else:
rot_pct = 1.0

n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
rope_dims = n_embd // n_head
rope_dims = int(rot_pct * n_embd) // n_head

# write rope scaling for long context (128k) model
rope_scaling = self.find_hparam(['rope_scaling'], True)
Expand Down Expand Up @@ -2565,7 +2582,7 @@ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')

if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}. long_factors = {len(long_factors)}, short_factors = {len(short_factors)}.')

yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
Expand Down
1 change: 1 addition & 0 deletions convert_hf_to_gguf_update.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,6 +109,7 @@ class TOKENIZER_TYPE(IntEnum):
{"name": "megrez", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
{"name": "deepseek-v3", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
{"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
{"name": "gpt-4o", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this does not actually work since Xenova/gpt-4o misses config.json, I have to make an exception for it

]


Expand Down
1 change: 1 addition & 0 deletions include/llama.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,6 +105,7 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26,
LLAMA_VOCAB_PRE_TYPE_MINERVA = 27,
LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28,
LLAMA_VOCAB_PRE_TYPE_GPT4O = 29,
};

enum llama_rope_type {
Expand Down
18 changes: 15 additions & 3 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2208,7 +2208,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {

// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);

// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}

for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
Expand All @@ -2223,8 +2228,15 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }, 0);

layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this check works but not actually correct, since scaling_type is to calculate attn_factor

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, the else branch is redundant because we know from the conversion script that if rot_pct is not set, then we're sure that n_rot = n_embd / n_head

Other arch like LLM_ARCH_LLAMA does the same thing

layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
}
else
{
layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
}
}
} break;
case LLM_ARCH_PHIMOE:
Expand Down
11 changes: 11 additions & 0 deletions src/llama-vocab.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -392,6 +392,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
};
break;
case LLAMA_VOCAB_PRE_TYPE_GPT4O:
// original regex from tokenizer.json
// "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
regex_exprs = {
"[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
default:
// default regex for BPE tokenization pre-processing
regex_exprs = {
Expand Down Expand Up @@ -1592,6 +1599,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
} else if (
tokenizer_pre == "megrez") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
} else if (
tokenizer_pre == "gpt-4o") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
clean_spaces = false;
} else {
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
}
Expand Down