[Speculative Decoding][Cherry Pick]Update extract_mtp_weight script and optimize config (#5213)

* update extract_mtp_model

* modify config usage

fastdeploy/config.py

@@ -1628,10 +1628,6 @@ class FDConfig:
                 else:
                     self.scheduler_config.max_num_batched_tokens = self.model_config.max_model_len
 
-        self.scheduler_config.max_chunk_len = (
-            self.scheduler_config.max_num_batched_tokens + self.scheduler_config.max_extra_num_batched_tokens
-        )
-
         if self.long_prefill_token_threshold == 0:
             self.long_prefill_token_threshold = int(self.model_config.max_model_len * 0.04)
 

fastdeploy/scheduler/config.py

@@ -270,7 +270,6 @@ class SchedulerConfig:
         self.name = "local"  # "local" for LocalScheduler or "global" for GlobalScheduler
         self.max_num_batched_tokens = 2048  # base token_num for text inputs
         self.max_extra_num_batched_tokens = 16384  # extra token_num for multimodal inputs
-        self.max_chunk_len = 18432  # max supported token_num = max_num_batched_tokens + max_extra_num_batched_tokens
         self.max_num_seqs = 34
         self.splitwise_role = "mixed"
         self.config = None

fastdeploy/spec_decode/mtp.py

@@ -355,7 +355,13 @@ class MTPProposer(Proposer):
             self.target_model_inputs["decoder_tile_ids_per_batch"]
         )
         self.model_inputs["target_hidden_states"] = paddle.full(
-            [self.fd_config.scheduler_config.max_chunk_len, self.model_config.hidden_size], 0, dtype="bfloat16"
+            [
+                self.fd_config.scheduler_config.max_num_batched_tokens
+                + self.fd_config.scheduler_config.max_extra_num_batched_tokens,
+                self.model_config.hidden_size,
+            ],
+            0,
+            dtype="bfloat16",
         )
 
         tmp_position_ids = paddle.arange(self.model_config.max_model_len).reshape((1, -1))

scripts/extract_mtp_weight_from_safetensor.py

@@ -17,7 +17,9 @@
 import argparse
 import json
 import os
+import re
 
+import numpy as np
 import paddle
 from paddleformers.transformers.model_utils import shard_checkpoint
 from paddleformers.utils.env import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME
@@ -46,6 +48,28 @@ def parse_args():
     return parser.parse_args()
 
 
+def dtype_byte_size(dtype):
+    """
+    Returns the size (in bytes) occupied by one parameter of type `dtype`.
+
+    Example:
+
+    ```py
+    >>> dtype_byte_size(paddle.float32)
+    4
+    ```
+    """
+    if str(dtype) in {"paddle.bool", "bool"}:
+        return 1 / 8
+    if str(dtype) in {"paddle.float8_e4m3fn", "paddle.float8_e5m2", "float8_e4m3fn", "float8_e5m2"}:
+        return 1
+    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
+    if bit_search is None:
+        raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
+    bit_size = int(bit_search.groups()[0])
+    return bit_size // 8
+
+
 def extract_mtp_weights(input_dir: str) -> dict:
     """
     Load all MTP-related weights from safetensors files in input_dir.
@@ -103,6 +127,18 @@ def save_safetensors(state_dict: dict, output_dir: str):
         logger.info(f"Saving shard: {save_path}")
         safe_save_file(shard, save_path, metadata={"format": "np"})
 
+    # If only one shard is returned, SAFE_WEIGHTS_INDEX_NAME will be null
+    if len(shards) == 1:
+        logger.info("Generate index file for single shard")
+        weight_size = 0
+        for key, weight in shards["model.safetensors"].items():
+            weight_size += np.prod(weight.shape) * dtype_byte_size(weight.dtype)
+
+        index = {
+            "metadata": {"total_size": int(weight_size)},
+            "weight_map": {k: "model.safetensors" for k in shards["model.safetensors"].keys()},
+        }
+
     index_path = os.path.join(output_dir, SAFE_WEIGHTS_INDEX_NAME)
     with open(index_path, "w", encoding="utf-8") as f:
         json.dump(index, f, indent=2)