apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 00:33:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix top_p_candidates and support separate setting of sampling params for mtp (#4189)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details

Browse Source 
* fix top_p_candidates

* For separate setting params for mtp

* delete print

* fix
This commit is contained in:
GoldPancake
2025-09-28 11:41:20 +08:00
committed by GitHub
parent f8c6a354a1
commit 076c30cb0f
4 changed files with 29 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
custom_ops/gpu_ops/speculate_decoding/speculate_verify.cu
View File

@@ -38,7 +38,13 @@ __device__ int64_t topp_sampling_kernel(const int64_t *candidate_ids,
const int tid = threadIdx.x;
float sum_scores = 0.0f;
float rand_top_p = curand_uniform(dev_curand_states + tid) * topp;
for (int i = 0; i < candidate_len; i++) {
sum_scores += candidate_scores[i];
}
float tgt_topp = sum_scores < topp ? sum_scores : topp;
sum_scores = 0.0f;
float rand_top_p = curand_uniform(dev_curand_states + tid) * tgt_topp;
for (int i = 0; i < candidate_len; i++) {
sum_scores += candidate_scores[i];
if (rand_top_p <= sum_scores) {

3
custom_ops/gpu_ops/speculate_decoding/top_p_candidates.cu
View File

@@ -467,6 +467,9 @@ __global__ void KeMatrixTopPBeamTopKFt(
break;
}
}
if (top_p_value == 1.0 && actual_candidates_lens[token_id] == 0){
actual_candidates_lens[token_id] = max_cadidate_len;
}
}
}

7
fastdeploy/envs.py
View File

@@ -95,6 +95,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
"FD_FOR_TORCH_MODEL_FORMAT": lambda: bool(int(os.getenv("FD_FOR_TORCH_MODEL_FORMAT", "0"))),
# force disable default chunked prefill
"FD_DISABLE_CHUNKED_PREFILL": lambda: bool(int(os.getenv("FD_DISABLE_CHUNKED_PREFILL", "0"))),
# For separate setting of sampling parameters for speculative decoding
"FD_SPECULATE_SAMPLING_TOP_P": lambda: (
None if "FD_SPECULATE_SAMPLING_TOP_P" not in os.environ else float(os.environ["FD_SPECULATE_SAMPLING_TOP_P"])
),
"FD_SPECULATE_SAMPLING_TOP_K": lambda: (
None if "FD_SPECULATE_SAMPLING_TOP_K" not in os.environ else float(os.environ["FD_SPECULATE_SAMPLING_TOP_K"])
),
"FD_ENABLE_INTERNAL_ADAPTER": lambda: int(os.getenv("FD_ENABLE_INTERNAL_ADAPTER", "0")),
# LLMEngine recieve requests port, used when FD_ENABLE_INTERNAL_ADAPTER=1
"FD_ZMQ_RECV_REQUEST_SERVER_PORT": lambda: os.getenv("FD_ZMQ_RECV_REQUEST_SERVER_PORT", "8200"),

12
fastdeploy/spec_decode/mtp.py
View File

@@ -303,8 +303,16 @@ class MTPProposer(Proposer):
)
# self.model_inputs["caches"] = self.cache_kvs
# Inherit generation hyperparameters from the main model for consistency
self.model_inputs["top_p"] = self.target_model_inputs["top_p"]
self.model_inputs["top_k"] = self.target_model_inputs["top_k"]
self.model_inputs["top_p"] = (
self.target_model_inputs["top_p"]
if envs.FD_SPECULATE_SAMPLING_TOP_P is None
else paddle.full_like(self.target_model_inputs["top_p"], envs.FD_SPECULATE_SAMPLING_TOP_P)
)
self.model_inputs["top_k"] = (
self.target_model_inputs["top_k"]
if envs.FD_SPECULATE_SAMPLING_TOP_K is None
else paddle.full_like(self.target_model_inputs["top_k"], envs.FD_SPECULATE_SAMPLING_TOP_K)
)
self.model_inputs["temperature"] = self.target_model_inputs["temperature"]
self.model_inputs["eos_token_id"] = self.target_model_inputs["eos_token_id"]
self.model_inputs["penalty_score"] = self.target_model_inputs["penalty_score"]