Skip to content

Add support for Microsoft Phi-4 model #10817

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

Merged
fairydreaming merged 4 commits into from
Dec 19, 2024
Merged

Add support for Microsoft Phi-4 model #10817

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
7555ab1
convert-hf : use GPT2 vocab and ignore sliding_window hparam for Phi-…
sszymczy Dec 13, 2024
520e8a0
convert-hf : do not use model name to distinguish Phi-4 from Phi-3
sszymczy Dec 14, 2024
c7fdbd3
convert-hf : use zero value of sliding_window to distinguish Phi-4 fr…
sszymczy Dec 14, 2024
046c0d7
llama : use zero value of n_swa to distinguish Phi-4 from other PHI3 …
sszymczy Dec 14, 2024
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
6 changes: 5 additions & 1 deletion convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2129,6 +2129,9 @@ class Phi3MiniModel(Model):
model_arch = gguf.MODEL_ARCH.PHI3

def set_vocab(self):
if self.metadata.name == "Phi 4":
return self._set_vocab_gpt2()
Copy link
Collaborator

@compilade compilade Dec 14, 2024

Choose a reason for hiding this comment

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

Alternatively, self._set_vocab_gpt2() could be called when tokenizer.model is missing here, regardless of the model name.

Copy link
Collaborator Author

@fairydreaming fairydreaming Dec 14, 2024

Choose a reason for hiding this comment

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

I modified the solution to check value of tokenizer_class from tokenizer_config.json and call self._set_vocab_gpt2() if it's GPT2Tokenizer.


from sentencepiece import SentencePieceProcessor

tokenizer_path = self.dir_model / 'tokenizer.model'
Expand Down Expand Up @@ -2245,7 +2248,8 @@ def set_gguf_parameters(self):
self.gguf_writer.add_rope_dimension_count(rope_dims)
self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
if self.metadata.name != "Phi 4":
self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))

def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
n_embd = self.find_hparam(["hidden_size", "n_embd"])
Expand Down
12 changes: 9 additions & 3 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5807,7 +5807,7 @@ static void llm_load_hparams(
hparams.n_swa = 131072;
}
bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
if (!found_swa && hparams.n_swa == 0) {
if (!found_swa && hparams.n_swa == 0 && model.name != "Phi 4") {
throw std::runtime_error("invalid value for sliding_window");
}
} break;
Expand Down Expand Up @@ -12839,7 +12839,13 @@ struct llm_build_context {
struct ggml_tensor * inp_pos = build_inp_pos();

// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
struct ggml_tensor * KQ_mask = nullptr;
if (model.name == "Phi 4") {
Copy link
Member

@slaren slaren Dec 14, 2024

Choose a reason for hiding this comment

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

I think a better solution would be to check if hparams.n_swa != 0.

Copy link
Collaborator Author

@fairydreaming fairydreaming Dec 14, 2024

Choose a reason for hiding this comment

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

I modified my patch to explicitly store zero sliding_window in case it's null in config.json and use the zero value to distinguish Phi-4 from other PHI3-based models.

// Phi-4 doesn't use sliding window attention
KQ_mask = build_inp_KQ_mask();
} else {
KQ_mask = build_inp_KQ_mask_swa();
}

for (int il = 0; il < n_layer; ++il) {
auto residual = inpL;
Expand Down Expand Up @@ -12897,7 +12903,7 @@ struct llm_build_context {

cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}

if (il == n_layer - 1) {
Expand Down
Loading