[OPs] Universal optimization and Fix early_stop cuda 700 (#3375)

* delete nonzero

* delete setup_ops_base.py

* check if

* check gcp infer_seed.cpu()

* fix repetition_early_stopper_kernel cuda 700

fastdeploy/model_executor/layers/sample/early_stopper.py

@@ -90,10 +90,10 @@ class RepetitionEarlyStopper(EarlyStopper):
         )
 
         B, W = self.trunc_scores.shape
-        V = probs.shape[1]
+        real_bsz, V = probs.shape
         BLOCK_W = triton.next_power_of_2(W)
 
-        grid = (B,)
+        grid = (real_bsz,)
         repetition_early_stopper_kernel[grid](
             self.trunc_scores,
             probs,

fastdeploy/model_executor/layers/sample/meta_data.py

@@ -42,7 +42,9 @@ class SamplingMetadata:
 
     top_p: paddle.Tensor
     top_k: Optional[paddle.Tensor] = None
+    top_k_list: Optional[list] = None
     min_p: Optional[paddle.Tensor] = None
+    min_p_list: Optional[list] = None
     seed: Optional[paddle.Tensor] = None
     max_num_logprobs: Optional[int] = None
     enable_early_stop: Optional[int] = False

fastdeploy/model_executor/layers/sample/ops/top_k_top_p_sampling.py

@@ -29,6 +29,7 @@ def top_k_top_p_sampling(
     x: paddle.Tensor,
     top_p: paddle.Tensor,
     top_k: Optional[paddle.Tensor] = None,
+    top_k_list: Optional[list] = None,
     threshold: Optional[paddle.Tensor] = None,
     topp_seed: Optional[paddle.Tensor] = None,
     seed: int = -1,
@@ -64,7 +65,7 @@ def top_k_top_p_sampling(
     if top_p_class == "air":
         _, ids = air_top_p_sampling(x, top_p, threshold, topp_seed, seed=seed, k=k, mode=mode)
     elif top_p_class == "rejection":
-        ids = rejection_top_p_sampling(x, top_p, top_k, seed, order)
+        ids = rejection_top_p_sampling(x, top_p, top_k, top_k_list, seed, order)
         _ = None
     elif top_p_class == "base_non_truncated":
         _, ids = paddle.tensor.top_p_sampling(
@@ -121,6 +122,7 @@ def rejection_top_p_sampling(
     x: paddle.Tensor,
     top_p: paddle.Tensor,
     top_k: paddle.Tensor,
+    top_k_list: list,
     seed: int = -1,
     order: Literal["top_k_first", "joint"] = "top_k_first",
 ) -> paddle.Tensor:
@@ -139,7 +141,7 @@ def rejection_top_p_sampling(
                 top_k_renorm_probs,
             )
 
-        if paddle.count_nonzero(top_k) == 0:
+        if not any(x > 0 for x in top_k_list):
             ids = rejection_top_p_sampling(
                 x,
                 top_p,
@@ -170,11 +172,12 @@ def rejection_top_p_sampling(
 def min_p_sampling(
     probs: paddle.tensor,
     min_p_arr: Optional[paddle.Tensor],
+    min_p_arr_cpu: Optional[list],
 ) -> tuple[paddle.Tensor, paddle.Tensor]:
     """
     min_p_sampling
     """
-    if paddle.count_nonzero(min_p_arr) == 0:
+    if not any(x > 0 for x in min_p_arr_cpu):
         return probs
     else:
         if current_platform.is_cuda():

fastdeploy/model_executor/layers/sample/sampler.py

@@ -281,10 +281,13 @@ class Sampler(nn.Layer):
 
         probs = F.softmax(logits)
 
-        probs = min_p_sampling(probs, sampling_metadata.min_p)
-
+        probs = min_p_sampling(probs, sampling_metadata.min_p, sampling_metadata.min_p_list)
         _, next_tokens = top_k_top_p_sampling(
-            probs, sampling_metadata.top_p, sampling_metadata.top_k, seed=sampling_metadata.seed[0, 0]
+            probs,
+            sampling_metadata.top_p,
+            sampling_metadata.top_k,
+            sampling_metadata.top_k_list,
+            seed=sampling_metadata.seed[0, 0],
         )
 
         logprobs_tensors = (

fastdeploy/model_executor/ops/gpu/__init__.py

@@ -19,7 +19,6 @@ from fastdeploy.import_ops import import_custom_ops
 
 PACKAGE = "fastdeploy.model_executor.ops.gpu"
 
-import_custom_ops(PACKAGE, "..base.fastdeploy_base_ops", globals())
 import_custom_ops(PACKAGE, ".fastdeploy_ops", globals())
 
 

fastdeploy/model_executor/ops/iluvatar/__init__.py

@@ -17,7 +17,6 @@ from fastdeploy.import_ops import import_custom_ops
 
 PACKAGE = "fastdeploy.model_executor.ops.iluvatar"
 
-import_custom_ops(PACKAGE, "..base.fastdeploy_base_ops", globals())
 import_custom_ops(PACKAGE, ".fastdeploy_ops", globals())
 
 from .moe_ops import iluvatar_moe_expert_ffn as moe_expert_ffn  # noqa: F401