feat: replace GeluAndMul (#1234)

python/sglang/srt/layers/activation.py

@@ -18,7 +18,7 @@ from typing import Optional
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from flashinfer.activation import gelu_tanh_and_mul, silu_and_mul
+from flashinfer.activation import gelu_and_mul, gelu_tanh_and_mul, silu_and_mul
 from vllm.distributed import (
     divide,
     get_tensor_model_parallel_rank,
@@ -43,18 +43,24 @@ class SiluAndMul(CustomOp):
 
 
 class GeluAndMul(CustomOp):
-    def __init__(self, **kwargs):
+    def __init__(self, approximate="tanh"):
         super().__init__()
+        self.approximate = approximate
 
     def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         d = x.shape[-1] // 2
-        return F.gelu(x[..., :d], approximate="tanh") * x[..., d:]
+        return F.gelu(x[..., :d], approximate=self.approximate) * x[..., d:]
 
     def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
         d = x.shape[-1] // 2
         output_shape = x.shape[:-1] + (d,)
         out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
-        gelu_tanh_and_mul(x, out)
+        if self.approximate == "tanh":
+            gelu_tanh_and_mul(x, out)
+        elif self.approximate == "none":
+            gelu_and_mul(x, out)
+        else:
+            raise RuntimeError("GeluAndMul only support tanh or none")
         return out
 
 

python/sglang/srt/models/gemma.py

@@ -23,7 +23,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.activation import GeluAndMul
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
     QKVParallelLinear,
@@ -34,6 +33,7 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -60,7 +60,7 @@ class GemmaMLP(nn.Module):
             bias=False,
             quant_config=quant_config,
         )
-        self.act_fn = GeluAndMul()
+        self.act_fn = GeluAndMul("none")
 
     def forward(self, x):
         gate_up, _ = self.gate_up_proj(x)

test/srt/models/test_generation_models.py

@@ -96,7 +96,7 @@ class TestGenerationModels(unittest.TestCase):
             if hf_logprobs.shape[0] <= 100:
                 assert torch.all(
                     abs(hf_logprobs - srt_logprobs) < prefill_tolerance
-                ), "prefill logprobs are not all close"
+                ), f"prefill logprobs are not all close with model_path={model_path} prompts={prompts} prefill_tolerance={prefill_tolerance}"
 
         print(f"hf_outputs.output_strs={hf_outputs.output_strs}")
         print(f"srt_outputs.output_strs={srt_outputs.output_strs}")