apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 00:57:33 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[BugFix] support real batch_size (#3109)

Browse Source 
* support real bsz

* fix

* fix xpu_model_runner.py,gpu_model_runner.py,gcu_model_runner.py,iluvatar_model_runner.py

* add event_loop_ep

* fix

* Add comments

* fix

* support mtp real_batch_size

* fix

* self.tmp_seq_lens_this_time->self.seq_lens_this_time_buffer

* fix

* fix VL real_seq_lens_this_time

* fix

* fix mtp

* fix

* fix mtp

* fix xpu

* fix
This commit is contained in:
lizexu123
2025-08-05 16:33:54 +08:00
committed by GitHub
parent 55939f7942
commit b01cfd6007
10 changed files with 110 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
fastdeploy/worker/gpu_model_runner.py
View File

@@ -164,6 +164,7 @@ class GPUModelRunner(ModelRunnerBase):
if self.speculative_method == "ngram":
self.proposer = NgramProposer(self.fd_config)
elif self.speculative_method == "mtp":
self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer
self.proposer = MTPProposer(
self.fd_config,
self.get_model(),
@@ -193,9 +194,11 @@ class GPUModelRunner(ModelRunnerBase):
return self.guided_backend.get_logits_processor(schemata_key=schemata_key), schemata_key
def insert_tasks_v1(self, req_dicts: List[Request]):
def insert_tasks_v1(self, req_dicts: List[Request], num_running_requests: int = None):
"""
Process scheduler output tasks, used when ENABLE_V1_KVCACHE_SCHEDULER=1
req_dict: A list of Request dict
num_running_requests: batch_size
"""
# NOTE(luotingdan): Lazy initialize kv cache
if "caches" not in self.share_inputs:
@@ -264,7 +267,7 @@ class GPUModelRunner(ModelRunnerBase):
)
self.share_inputs["stop_flags"][idx : idx + 1] = False
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = prefill_start_index
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = length
self.seq_lens_this_time_buffer[idx : idx + 1] = length
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = length
self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = 0
self.share_inputs["prompt_lens"][idx : idx + 1] = len(input_ids)
@@ -286,7 +289,7 @@ class GPUModelRunner(ModelRunnerBase):
logger.debug(f"Handle preempted request {request} at idx {idx}")
self.share_inputs["block_tables"][idx : idx + 1, :] = -1
self.share_inputs["stop_flags"][idx : idx + 1] = True
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = 0
self.seq_lens_this_time_buffer[idx : idx + 1] = 0
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = 0
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = 0
self.share_inputs["is_block_step"][idx : idx + 1] = False
@@ -328,10 +331,13 @@ class GPUModelRunner(ModelRunnerBase):
if has_prefill_task:
self.share_inputs["not_need_stop"][0] = True
self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer[:num_running_requests]
def insert_prefill_inputs(self, req_dicts: List[Request]):
def insert_prefill_inputs(self, req_dicts: List[Request], num_running_requests: int = None):
"""
Process inputs for prefill tasks and insert it to share_inputs buffer
req_dict: A list of Request dict
num_running_requests: batch_size
TODO(gongshaotian): Refactor this func
"""
@@ -365,7 +371,7 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["prompt_ids"][idx : idx + 1, :length] = np.array(request.prompt_token_ids)
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = 0
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = length
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = 1
self.seq_lens_this_time_buffer[idx : idx + 1] = 1
self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = 0
self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = length
self.share_inputs["prompt_lens"][idx : idx + 1] = length
@@ -377,7 +383,7 @@ class GPUModelRunner(ModelRunnerBase):
request.draft_token_ids[0:num_prefill_send_token],
dtype="int64",
)
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = num_prefill_send_token
self.seq_lens_this_time_buffer[idx : idx + 1] = num_prefill_send_token
else:
self.share_inputs["pre_ids"][idx : idx + 1] = -1
self.share_inputs["step_idx"][idx : idx + 1] = 0
@@ -412,7 +418,7 @@ class GPUModelRunner(ModelRunnerBase):
)
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = token_chunk_size
self.seq_lens_this_time_buffer[idx : idx + 1] = token_chunk_size
self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = token_chunk_size
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = token_chunk_size
self.share_inputs["prompt_lens"][idx : idx + 1] = token_chunk_size
@@ -430,7 +436,7 @@ class GPUModelRunner(ModelRunnerBase):
else:
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = length
self.seq_lens_this_time_buffer[idx : idx + 1] = length
self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = length
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = length
self.share_inputs["prompt_lens"][idx : idx + 1] = length
@@ -516,8 +522,10 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["not_need_stop"][0] = True
self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer[:num_running_requests]
if self.speculative_method in ["mtp"]:
self.proposer.insert_prefill_inputs(req_dicts)
self.proposer.insert_prefill_inputs(req_dicts, num_running_requests)
def _dummy_prefill_inputs(self, num_tokens: int, batch_size: int, expected_decode_len: int):
"""Set dummy prefill inputs to share_inputs"""
@@ -543,7 +551,7 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["input_ids"][idx : idx + 1, :input_length] = np.array([5] * input_length)
self.share_inputs["prompt_ids"][idx : idx + 1, :input_length] = np.array([5] * input_length)
self.share_inputs["eos_token_id"][:] = np.array([2], dtype="int64").reshape(-1, 1)
self.share_inputs["seq_lens_this_time"][idx : idx + 1] = input_length
self.seq_lens_this_time_buffer[idx : idx + 1] = input_length
self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = input_length
self.share_inputs["seq_lens_encoder"][idx : idx + 1] = input_length
self.share_inputs["seq_lens_decoder"][idx : idx + 1] = 0
@@ -561,6 +569,7 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["block_tables"][idx : idx + 1, :block_num] = np.arange(
idx * block_num, (idx + 1) * block_num, 1
)
self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer
def _init_share_inputs(self, max_num_seqs: int):
"""
@@ -611,7 +620,7 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["max_length"] = paddle.full(
[max_num_seqs, 1], self.model_config.max_model_len, dtype="int64"
)
self.share_inputs["seq_lens_this_time"] = paddle.full(max_num_seqs, 0, dtype="int32")
self.seq_lens_this_time_buffer = paddle.full(max_num_seqs, 0, dtype="int32")
self.share_inputs["seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
self.share_inputs["seq_lens_decoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
self.share_inputs["step_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
@@ -1255,6 +1264,7 @@ class GPUModelRunner(ModelRunnerBase):
def execute_model(
self,
model_forward_batch: Optional[List[Request]] = None,
num_running_requests: int = None,
) -> Optional[ModelRunnerOutput]:
"""
The Entrance of model execute.
@@ -1263,6 +1273,7 @@ class GPUModelRunner(ModelRunnerBase):
class at the server level, which is too granular for ModelRunner.
We plan to replace it with 'ModelForwardBatch'.
intermediate_tensors:
num_running_requests: batch_size
"""
# 1. Prepare inputs of model and sampler.
skip_idx_list = self._get_skip_idx(model_forward_batch)
@@ -1364,8 +1375,8 @@ class GPUModelRunner(ModelRunnerBase):
accept_num=(self.share_inputs["accept_num"] if self.speculative_decoding else None),
enable_thinking=(self.share_inputs["enable_thinking"] if self.enable_mm else None),
think_end_id=(self.model_config.think_end_id if self.enable_mm else -1),
need_think_end=(self.share_inputs["need_think_end"] if self.enable_mm else None),
reasoning_index=(self.share_inputs["reasoning_index"] if self.enable_mm else None),
need_think_end=(self.share_inputs["need_think_end"][:num_running_requests] if self.enable_mm else None),
reasoning_index=(self.share_inputs["reasoning_index"][:num_running_requests] if self.enable_mm else None),
stop_token_ids=self.share_inputs["stop_seqs"],
stop_seqs_len=self.share_inputs["stop_seqs_len"],
)
@@ -1405,6 +1416,10 @@ class GPUModelRunner(ModelRunnerBase):
self._update_chunked_prefill(model_forward_batch)
self._add_cache(model_forward_batch)
self.seq_lens_this_time_buffer[:num_running_requests].copy_(
self.share_inputs["seq_lens_this_time"][:num_running_requests], False
)
return None
def _add_cache(self, model_forward_batch) -> None: