Add DeepSeek R1 Distill 8B

tokenizer/hf_tokenizer.py

@@ -46,8 +46,14 @@ def __init__(self, file_path: str):
         if tokenizer_config_path is not None:
             with open(tokenizer_config_path, "r") as handle:
                 tok_config = json.load(handle)
-            bos_token = tok_config.get("bos_token")
-            eos_token = tok_config.get("eos_token")
+
+            def _extract_token(identifier: str) -> Optional[str]:
+                entry: Optional[Union[str, dict]] = tok_config.get(identifier)
+                return entry.get("content") if isinstance(entry, dict) else entry
+
+            bos_token = _extract_token("bos_token")
+            eos_token = _extract_token("eos_token")
+
             if bos_token is not None:
                 self._bos_id = self._tokenizer.token_to_id(bos_token)
             if eos_token is not None:

torchchat/model_config/models.json

@@ -51,6 +51,12 @@
         "distribution_path": "meta-llama/Meta-Llama-3.1-8B-Instruct",
         "transformer_params_key": "Meta-Llama-3.1-8B"
     },
+    "deepseek-ai/DeepSeek-R1-Distill-Llama-8B": {
+        "aliases": ["deepseek-r1-distill", "deepseek-r1-distill-8b"],
+        "distribution_channel": "HuggingFaceSnapshot",
+        "distribution_path": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
+        "tokenizer_file": "tokenizer.json"
+    },
     "meta-llama/Meta-Llama-3.1-70B-Instruct": {
         "aliases": ["llama3.1-70b"],
         "distribution_channel": "HuggingFaceSnapshot",

torchchat/model_params/DeepSeek-R1-Distill-Llama-8B.json

@@ -0,0 +1 @@
+{"block_size": 131072, "dim": 4096, "ffn_dim_multiplier": 1.3, "multiple_of": 1024, "n_heads": 32, "n_local_heads": 8, "n_layers": 32, "rope_base": 500000.0, "vocab_size": 128256, "use_tiktoken": true, "use_hf_tokenizer": true, "norm_eps": 1e-05, "rope_scaling": {"factor": 8.0, "low_freq_factor": 1.0, "high_freq_factor": 4.0, "original_max_position_embeddings": 8192}}