[New Model] support Baichuan-M1

Signed-off-by: dangshunya <[email protected]>

Author: dangshunya

docs/source/models/supported_models.md

@@ -96,6 +96,11 @@ See [this page](#generative-models) for more information on how to use generativ
   - `baichuan-inc/Baichuan2-13B-Chat`, `baichuan-inc/Baichuan-7B`, etc.
   - ✅︎
   - ✅︎
+* - `BaichuanM1ForCausalLM`
+  - Baichuan-M1
+  - `baichuan-inc/Baichuan-M1-14B-Instruct`, `baichuan-inc/Baichuan-M1-14B-Base`, etc.
+  - ✅︎
+  - ✅︎
 * - `BloomForCausalLM`
   - BLOOM, BLOOMZ, BLOOMChat
   - `bigscience/bloom`, `bigscience/bloomz`, etc.

tests/models/registry.py

@@ -100,6 +100,8 @@ def check_available_online(
                                          trust_remote_code=True),
     "BaichuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan2-7B-chat",
                                          trust_remote_code=True),
+    "BaichuanM1ForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan-M1-14B-Instruct", # noqa: E501
+                                         trust_remote_code=True),
     "BloomForCausalLM": _HfExamplesInfo("bigscience/bloomz-1b1"),
     # ChatGLMModel supports multimodal
     "CohereForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r-v01",

vllm/config.py

@@ -305,9 +305,12 @@ def __init__(
             self.enforce_eager = False
 
         sliding_window = getattr(self.hf_text_config, "sliding_window", None)
+        sliding_window_layers = getattr(self.hf_text_config,
+                                        "sliding_window_layers", None)
         has_interleaved_attention = (sliding_window is not None) and (
             isinstance(sliding_window, list) or
-            (self.hf_text_config.model_type in ["gemma2", "cohere2"]))
+            (self.hf_text_config.model_type in ["gemma2", "cohere2"])
+            or sliding_window_layers is not None)
 
         if (not self.disable_sliding_window and has_interleaved_attention):
             if envs.VLLM_ATTENTION_BACKEND == "XFORMERS":
@@ -713,6 +716,9 @@ def get_hf_config_sliding_window(
         if (hasattr(self.hf_text_config, "use_sliding_window")
                 and not self.hf_text_config.use_sliding_window):
             return None
+        if hasattr(self.hf_text_config, 'sliding_window_layers'):
+            return None
+
         return getattr(self.hf_text_config, "sliding_window", None)
 
     def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
@@ -724,6 +730,10 @@ def get_sliding_window(self) -> Optional[Union[int, List[Optional[int]]]]:
         # Otherwise get the value from the hf config.
         return self.get_hf_config_sliding_window()
 
+    def get_sliding_window_layers(self,
+                                  parallel_config) -> Optional[List[int]]:
+        return getattr(self.hf_text_config, "sliding_window_layers", [])
+
     def get_vocab_size(self) -> int:
         return self.hf_text_config.vocab_size
 
@@ -751,6 +761,12 @@ def get_head_size(self) -> int:
         return (self.hf_text_config.hidden_size //
                 self.hf_text_config.num_attention_heads)
 
+    def get_head_size_swa(self) -> int:
+        if hasattr(self.hf_text_config, "num_swa_attention_heads"):
+            return (self.hf_text_config.hidden_size //
+                    self.hf_text_config.num_swa_attention_heads)
+        return self.get_head_size()
+
     def get_total_num_kv_heads(self) -> int:
         """Returns the total number of KV heads."""
         # For GPTBigCode & Falcon:
@@ -797,6 +813,22 @@ def get_total_num_kv_heads(self) -> int:
         # equal to the number of attention heads.
         return self.hf_text_config.num_attention_heads
 
+    def get_total_num_kv_heads_swa(self) -> int:
+        if hasattr(self.hf_text_config, "num_swa_key_value_heads"):
+            return self.hf_text_config.num_swa_key_value_heads
+        return self.get_total_num_kv_heads()
+
+    def get_num_swa_key_value_heads(self,
+                                    parallel_config: "ParallelConfig") -> int:
+        """Returns the number of KV heads per GPU."""
+        total_num_kv_heads_swa = self.get_total_num_kv_heads_swa()
+        # If tensor parallelism is used, we divide the number of KV heads by
+        # the tensor parallel size. We will replicate the KV heads in the
+        # case where the number of KV heads is smaller than the tensor
+        # parallel size so each GPU has at least one KV head.
+        return max(
+            1, total_num_kv_heads_swa // parallel_config.tensor_parallel_size)
+
     def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
         """Returns the number of KV heads per GPU."""
         total_num_kv_heads = self.get_total_num_kv_heads()
@@ -839,7 +871,18 @@ def get_num_layers_by_block_type(
 
         if is_transformer:
             # Handle the basic case first
-            return end - start if attn_block_type else 0
+            swa_layers = self.get_sliding_window_layers(parallel_config)
+            num_layers = 0
+            if not swa_layers:
+                num_layers = end - start if attn_block_type else 0
+            else:
+                for layer_id in range(start, end):
+                    if (block_type == LayerBlockType.attention
+                            and layer_id not in swa_layers) or (
+                                block_type == LayerBlockType.swa
+                                and layer_id in swa_layers):
+                        num_layers += 1
+            return num_layers
         elif self.is_attention_free:
             # Attention free
             # Note that this code assumes there
@@ -2360,7 +2403,6 @@ def _get_and_verify_max_len(
             max_len_key = key if max_len < derived_max_model_len \
                 else max_len_key
             derived_max_model_len = min(derived_max_model_len, max_len)
-
     # If sliding window is manually disabled, max_length should be less
     # than the sliding window length in the model config.
     if disable_sliding_window and sliding_window_len is not None:

vllm/engine/llm_engine.py

@@ -341,6 +341,11 @@ def get_tokenizer_for_seq(sequence: Sequence) -> AnyTokenizer:
         # of request outputs to asyncio queues
         self.process_request_outputs_callback: Optional[Callable] = None
 
+        # This is used to evict the finished requests from the Mamba cache and
+        # Baichuan-M1, We should use it to keep finished_req_ids when scheduler
+        # is empty.
+        self.finished_requests_ids: List[str] = list()
+
         # Create the scheduler.
         # NOTE: the cache_config here have been updated with the numbers of
         # GPU and CPU blocks, which are profiled in the distributed executor.
@@ -1323,6 +1328,7 @@ def step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
 
             finished_requests_ids = self.scheduler[
                 virtual_engine].get_and_reset_finished_requests_ids()
+            self.finished_requests_ids.extend(finished_requests_ids)
 
             # Maybe switch from async mode to sync mode
             if not allow_async_output_proc and len(ctx.output_queue) > 0:
@@ -1335,8 +1341,6 @@ def step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
                 self._cache_scheduler_outputs_for_multi_step(
                     virtual_engine, seq_group_metadata_list, scheduler_outputs,
                     allow_async_output_proc)
-        else:
-            finished_requests_ids = list()
 
         assert seq_group_metadata_list is not None
         assert scheduler_outputs is not None
@@ -1357,11 +1361,14 @@ def step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
                 blocks_to_copy=scheduler_outputs.blocks_to_copy,
                 num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
                 running_queue_size=scheduler_outputs.running_queue_size,
-                finished_requests_ids=finished_requests_ids,
+                finished_requests_ids=self.finished_requests_ids,
                 # We use ExecuteModelRequest to pass the last sampled_token_ids
                 # to each of the non-last PP stages for in-place prepare_input.
                 last_sampled_token_ids=last_sampled_token_ids)
 
+            # Clear finished_requests_ids list.
+            self.finished_requests_ids = list()
+
             if allow_async_output_proc:
                 execute_model_req.async_callback = self.async_callbacks[
                     virtual_engine]