apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 00:06:38 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Excutor] Experiment Feature-Support Prefill in cudagraph (#3459)

Browse Source 
* Support prefill in Cudagraph

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2.1

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2.2

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2.3

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2.4

* Refactor GetBlockShapeAndSplitKVBlock Kernel V2.5

* Solve problem about encoder_num_blocks_x_cpu

* Add early-exit mechanism for attention kernel

* fix test case about append-attention

* Update testcode, Add annotations to related tensors

* move get_input_length_list

* solve test_code

* Add annotations about early-exit for attention kernel

* Add annotations about early-exit for attention kernel2

* solve comment

* solve mtp

---------

Co-authored-by: RAM <gstian5555@outlook.com>
This commit is contained in:
Jundong Liu
2025-09-08 13:12:24 +08:00
committed by GitHub
parent 472402bf4e
commit 3d0aaa5923
21 changed files with 528 additions and 260 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
fastdeploy/spec_decode/mtp.py
View File

@@ -212,6 +212,22 @@ class MTPProposer(Proposer):
self.target_model_inputs["max_len_tensor_cpu"]
).cpu()
self.model_inputs["encoder_batch_ids"] = paddle.zeros_like(self.target_model_inputs["encoder_batch_ids"])
self.model_inputs["encoder_tile_ids_per_batch"] = paddle.zeros_like(
self.target_model_inputs["encoder_tile_ids_per_batch"]
)
self.model_inputs["encoder_num_blocks_x_cpu"] = paddle.zeros_like(
self.target_model_inputs["encoder_num_blocks_x_cpu"]
).cpu()
self.model_inputs["kv_batch_ids"] = paddle.zeros_like(self.target_model_inputs["kv_batch_ids"])
self.model_inputs["kv_tile_ids_per_batch"] = paddle.zeros_like(
self.target_model_inputs["kv_tile_ids_per_batch"]
)
self.model_inputs["kv_num_blocks_x_cpu"] = paddle.zeros_like(
self.target_model_inputs["kv_num_blocks_x_cpu"]
).cpu()
self.model_inputs["max_len_kv_cpu"] = paddle.zeros_like(self.target_model_inputs["max_len_kv_cpu"]).cpu()
# Get the attention backend
attn_cls = get_attention_backend()
attn_backend = attn_cls(
@@ -321,6 +337,13 @@ class MTPProposer(Proposer):
self.model_inputs["decoder_tile_ids_per_batch"] = None
self.model_inputs["decoder_num_blocks_cpu"] = None # Pinning Memory
self.model_inputs["max_len_tensor_cpu"] = None # CPU
self.model_inputs["encoder_batch_ids"] = None
self.model_inputs["encoder_tile_ids_per_batch"] = None
self.model_inputs["encoder_num_blocks_x_cpu"] = None # CPU
self.model_inputs["kv_batch_ids"] = None
self.model_inputs["kv_tile_ids_per_batch"] = None
self.model_inputs["kv_num_blocks_x_cpu"] = None # CPU
self.model_inputs["max_len_kv_cpu"] = None # CPU
# Input tokens
self.model_inputs["draft_tokens"] = paddle.full(
@@ -512,6 +535,13 @@ class MTPProposer(Proposer):
cu_seqlens_k=self.model_inputs["cu_seqlens_k"],
block_tables=self.model_inputs["block_tables"],
caches=self.model_inputs["caches"],
encoder_batch_ids=self.model_inputs["encoder_batch_ids"],
encoder_tile_ids_per_batch=self.model_inputs["encoder_tile_ids_per_batch"],
encoder_num_blocks_x_cpu=self.model_inputs["encoder_num_blocks_x_cpu"],
kv_batch_ids=self.model_inputs["kv_batch_ids"],
kv_tile_ids_per_batch=self.model_inputs["kv_tile_ids_per_batch"],
kv_num_blocks_x_cpu=self.model_inputs["kv_num_blocks_x_cpu"],
max_len_kv_cpu=self.model_inputs["max_len_kv_cpu"],
)
# Initialzie attention meta data