convert : add RuGPT3XL (RuGPT3XLForCausalLM) support (#21011)

* Support of ruGPT3XL model added

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <redacted>
* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <redacted>
* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <redacted>
* chkhsh for ruGPT3XL model added

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <redacted>
* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <redacted>
* Fixing chkhsh for ruGPT3XL, rerun updated and _qkv_parts in RuGPT3XLModel

---------

Co-authored-by: Sigbjørn Skjæret <redacted>

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index f377738f8dd96fdf63315ce8ebc09c0061060057..1e2fce7edda856608eae7cf9bec2b16f9ae0ee58 100755 (executable)
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -1311,6 +1311,9 @@ class TextModel(ModelBase):
         if chkhsh == "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df":
             # ref: https://huggingface.co/aari1995/German_Semantic_V3
             res = "jina-v2-de"
+        if chkhsh == "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4":
+            # ref: https://huggingface.co/evilfreelancer/ruGPT3XL
+            res = "gpt-2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -5100,6 +5103,47 @@ class GPT2Model(TextModel):
         yield from super().modify_tensors(data_torch, new_name, bid)
 
 
+@ModelBase.register("RuGPT3XLForCausalLM")
+class RuGPT3XLModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.GPT2
+
+    _qkv_parts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Fuse separate Q, K, V projections into a single QKV tensor
+        if ".self_attn.q_proj." in name or ".self_attn.k_proj." in name or ".self_attn.v_proj." in name:
+            suffix = "weight" if name.endswith(".weight") else "bias"
+            part = "q" if ".q_proj." in name else ("k" if ".k_proj." in name else "v")
+            key = f"{part}.{suffix}"
+
+            assert bid is not None
+            if self._qkv_parts is None:
+                self._qkv_parts = [{} for _ in range(self.block_count)]
+            self._qkv_parts[bid][key] = data_torch
+
+            q_key, k_key, v_key = f"q.{suffix}", f"k.{suffix}", f"v.{suffix}"
+            if all(k in self._qkv_parts[bid] for k in [q_key, k_key, v_key]):
+                q = self._qkv_parts[bid].pop(q_key)
+                k = self._qkv_parts[bid].pop(k_key)
+                v = self._qkv_parts[bid].pop(v_key)
+                data_torch = torch.cat([q, k, v], dim=0)
+                name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid, f".{suffix}")
+                logger.debug(f"Fused Q/K/V {suffix} for layer {bid} -> {name}")
+            else:
+                return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._qkv_parts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            parts = [f"({i}){k}" for i, d in enumerate(self._qkv_parts) for k in d.keys()]
+            if len(parts) > 0:
+                raise ValueError(f"Unprocessed Q/K/V parts: {parts}")
+
+
 @ModelBase.register("PhiForCausalLM")
 class Phi2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI2

diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index 1e8b29fb2ec0cece34b934e70cbb44f470d1f27b..086f1c22863a6255b4414b7357c8aa69de4745c6 100755 (executable)
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -178,6 +178,7 @@ pre_computed_hashes = [
     {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
     # jina-v2-de variants
     {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
+    {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/evilfreelancer/ruGPT3XL", "chkhsh": "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4"},
 ]
 
 

diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 281c1a830484ed9962a24310bb0f11acfcd47c7b..df70577dbc2973ad7430a7bea324a2e4e94759ac 100644 (file)
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -63,6 +63,7 @@ class TensorNameMap:
             "transformer.wpe",                 # gpt2
             "embeddings.position_embeddings",  # bert
             "wpe",                             # gpt2
+            "model.embed_positions",           # rugpt3xl
         ),
 
         # Output