【Hackathon 9th No.93】Add Minimax-m1 for FastDeploy

fastdeploy/demo/offline_demo.py

@@ -17,11 +17,11 @@
 from fastdeploy.engine.sampling_params import SamplingParams
 from fastdeploy.entrypoints.llm import LLM
 
-model_name_or_path = "./models/llama-7b"
+model_name_or_path = "/home/aistudio/config_folder"
 
 # 超参设置
 sampling_params = SamplingParams(temperature=0.1, max_tokens=30)
-llm = LLM(model=model_name_or_path, tensor_parallel_size=1)
+llm = LLM(model=model_name_or_path, tensor_parallel_size=4, load_choices="default_v1")
 output = llm.generate(prompts="who are you？", use_tqdm=True, sampling_params=sampling_params)
 
 print(output)

fastdeploy/engine/engine.py

@@ -120,6 +120,32 @@ def start(self, api_server_pid=None):
         self._init_worker_signals()
 
         self.data_processor = self.input_processor.create_processor()
+
+        # [CORE FIX] Check and set pad_token_id if it's missing.
+        # Some tokenizers do not have a pad_token_id, which causes issues with padding.
+        # We use the eos_token_id as a robust fallback in such cases.
+        if self.data_processor.pad_token_id is None:
+            eos_token_id = self.data_processor.tokenizer.eos_token_id
+            if eos_token_id is not None:
+                console_logger.warning(
+                    f"Tokenizer's pad_token_id is None. Setting it to the eos_token_id ({eos_token_id}) for padding."
+                )
+                # 1. Update the tokenizer instance directly. This is crucial for padding operations.
+                self.data_processor.tokenizer.pad_token_id = eos_token_id
+
+                # 2. Update the data_processor's attribute to ensure workers are initialized with the correct value.
+                self.data_processor.pad_token_id = eos_token_id
+
+                # 3. Update the main model configuration to maintain a single source of truth.
+                if self.cfg.model_config.pad_token_id is None:
+                    self.cfg.model_config.pad_token_id = eos_token_id
+            else:
+                # This is a critical failure case. A model must have a token for padding.
+                raise ValueError(
+                    "Tokenizer has neither a pad_token_id nor an eos_token_id. "
+                    "Cannot proceed without a token for padding."
+                )
+
         self.engine.data_processor = self.data_processor
         # Launch components: scheduler, cache_manager, expert_service et.al.
         self.launch_components()

fastdeploy/model_executor/forward_meta.py

@@ -142,6 +142,11 @@ class ForwardMeta:
     block_tables: Optional[paddle.Tensor] = None
     # KV caches
     caches: Optional[list[paddle.Tensor]] = None
+
+    # Linear attention caches
+    linear_attn_caches: Optional[paddle.Tensor] = None
+    # Slot mapping
+    slot_mapping: Optional[paddle.Tensor] = None
 
     def clear_caches(self):
         """Safely clean up the caches"""

fastdeploy/model_executor/layers/rotary_embedding.py

@@ -276,7 +276,7 @@ def get_rope_impl(
     if model_config is None or architecture.startswith("Qwen"):
         rotary_emb_layer = QwenRotaryEmbedding(rotary_dim, base, partial_rotary_factor)
         rotary_emb = rotary_emb_layer(position_ids)
-    elif architecture.startswith("Glm"):
+    elif "MiniMaxM1" in architecture or architecture.startswith("Glm"):
         rotary_emb_layer = GlmRotaryEmbedding(rotary_dim, base, partial_rotary_factor)
         rotary_emb = rotary_emb_layer(position_ids)
     else: