[Feature] support min_p_sampling (#2872)

* Fastdeploy support min_p

* add test_min_p

* fix

* min_p_sampling

* update

* delete vl_gpu_model_runner.py

* fix

* Align usage of min_p with vLLM

* fix

* modified unit test

* fix test_min_sampling

* pre-commit all files

* fix

* fix

* fix

* fix xpu_model_runner.py

fastdeploy/model_executor/layers/sample/meta_data.py

@@ -42,6 +42,7 @@ class SamplingMetadata:
 
     top_p: paddle.Tensor
     top_k: Optional[paddle.Tensor] = None
+    min_p: Optional[paddle.Tensor] = None
     max_num_logprobs: Optional[int] = None
     prompt_ids: Optional[paddle.Tensor] = None
     prompt_lens: Optional[paddle.Tensor] = None

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

@@ -18,10 +18,11 @@ from .apply_penalty_multi_scores import (
     apply_penalty_multi_scores,
     apply_speculative_penalty_multi_scores,
 )
-from .top_k_top_p_sampling import top_k_top_p_sampling
+from .top_k_top_p_sampling import min_p_sampling, top_k_top_p_sampling
 
 __all__ = [
     "apply_penalty_multi_scores",
     "apply_speculative_penalty_multi_scores",
     "top_k_top_p_sampling",
+    "min_p_sampling",
 ]

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

@@ -60,6 +60,7 @@ def top_k_top_p_sampling(
 
     """
     top_p_class = envs.FD_SAMPLING_CLASS.lower()
+
     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":
@@ -154,3 +155,25 @@ def rejection_top_p_sampling(
     except ImportError:
         raise RuntimeError("Cannot import rejection_top_p_sampling op.")
     return ids
+
+
+def min_p_sampling(
+    probs: paddle.tensor,
+    min_p_arr: Optional[paddle.Tensor],
+) -> tuple[paddle.Tensor, paddle.Tensor]:
+    """
+    min_p_sampling
+    """
+    if paddle.count_nonzero(min_p_arr) == 0:
+        return probs
+    else:
+        if current_platform.is_cuda():
+            from fastdeploy.model_executor.ops.gpu import min_p_sampling
+
+            probs = min_p_sampling(probs, min_p_arr)
+        else:
+            max_probabilities = paddle.amax(probs, axis=-1, keepdim=True)
+            adjusted_min_p = max_probabilities * min_p_arr
+            invalid_token_mask = probs < adjusted_min_p.reshape([-1, 1])
+            probs = paddle.where(invalid_token_mask, paddle.full_like(probs, 0.0), probs)
+        return probs

fastdeploy/model_executor/layers/sample/sampler.py

@@ -30,6 +30,7 @@ from fastdeploy.model_executor.layers.sample.meta_data import SamplingMetadata
 from fastdeploy.model_executor.layers.sample.ops import (
     apply_penalty_multi_scores,
     apply_speculative_penalty_multi_scores,
+    min_p_sampling,
     top_k_top_p_sampling,
 )
 from fastdeploy.platforms import current_platform
@@ -266,6 +267,8 @@ class Sampler(nn.Layer):
 
         probs = F.softmax(logits)
 
+        probs = min_p_sampling(probs, sampling_metadata.min_p)
+
         _, next_tokens = top_k_top_p_sampling(probs, sampling_metadata.top_p, sampling_metadata.top_k)
 
         logprobs_tensors = (
@@ -281,6 +284,7 @@ class Sampler(nn.Layer):
             sampled_token_ids=next_tokens,
             logprobs_tensors=logprobs_tensors,
         )
+
         return sampler_output