remove xformers deps

README.md

@@ -19,10 +19,10 @@ vLLM是一个快速且易于使用的LLM推理和服务库,使用PageAttention
 | BloomForCausalLM      | BLOOM                                                       | Yes | Yes |
 | InternLMForCausalLM   | InternLM                                                    | Yes | Yes |
 | InternLM2ForCausalLM  | InternLM2                                                   | Yes | Yes |
+| TeleChat12BForCausalLM (#TelechatForCausalLM) | TeleChat-12B                        | Yes | Yes |
 | MiniCPMForCausalLM    | MiniCPM                                                     | Yes | Yes |
 | MiniCPM3ForCausalLM   | MiniCPM3                                                    | Yes | Yes |
 | MixtralForCausalLM    | Mixtral-8x7B,Mixtral-8x7B-Instruct                          | Yes | Yes |
-| TeleChat12BForCausalLM (#TelechatForCausalLM) | TeleChat-12B                        | Yes | Yes |
 | Qwen2MoeForCausalLM                 | Qwen2-57B-A14B,Qwen2-57B-A14B-Instruct        | Yes | Yes |
 | LlavaForConditionalGeneration       | LLaMA,LLaMA-2,LLaMA-3                         | Yes | Yes |
 | Qwen2VLForConditionalGeneration     | Qwen2-VL                                      | Yes | Yes |
@@ -74,7 +74,6 @@ VLLM_INSTALL_PUNICA_KERNELS=1 python3 setup.py install
 
 2、根据pytorch2.3.0、python、dtk及系统下载对应的依赖包:
 - triton:[https://cancon.hpccube.com:65024/4/main/triton](https://cancon.hpccube.com:65024/4/main/triton/)
-- xformers:[https://cancon.hpccube.com:65024/4/main/xformers](https://cancon.hpccube.com:65024/4/main/xformers)
 - flash_attn: [https://cancon.hpccube.com:65024/4/main/flash_attn](https://cancon.hpccube.com:65024/4/main/flash_attn)
 - lmslim: [https://cancon.hpccube.com:65024/4/main/lmslim](https://cancon.hpccube.com:65024/4/main/lmslim)
 

requirements-rocm.txt

@@ -14,5 +14,4 @@ setuptools_scm>=8
 torch == 2.3.0
 triton == 2.1.0
 flash_attn == 2.6.1
-xformers == 0.0.25
-lmslim == 0.1.2
\ No newline at end of file
+lmslim == 0.1.2  # future version 0.2.0
\ No newline at end of file

vllm/model_executor/models/blip.py

@@ -20,10 +20,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import SequenceData
+import vllm.envs as envs
 
 try:
-    from xformers import ops as xops
-    USE_XFORMERS_OPS = True
+    if envs.VLLM_ATTENTION_BACKEND=="XFormers":
+        from xformers import ops as xops
+        USE_XFORMERS_OPS = True
+    else:
+        USE_XFORMERS_OPS = False
 except ImportError:
     USE_XFORMERS_OPS = False
 

vllm/model_executor/models/clip.py

@@ -21,10 +21,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import SequenceData
+import vllm.envs as envs
 
 try:
-    from xformers import ops as xops
-    USE_XFORMERS_OPS = True
+    if envs.VLLM_ATTENTION_BACKEND=="XFormers":
+        from xformers import ops as xops
+        USE_XFORMERS_OPS = True
+    else:
+        USE_XFORMERS_OPS = False
 except ImportError:
     USE_XFORMERS_OPS = False
 

vllm/model_executor/models/intern_vit.py

@@ -19,10 +19,14 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+import vllm.envs as envs
 
 try:
-    from xformers import ops as xops
-    USE_XFORMERS_OPS = True
+    if envs.VLLM_ATTENTION_BACKEND=="XFormers":
+        from xformers import ops as xops
+        USE_XFORMERS_OPS = True
+    else:
+        USE_XFORMERS_OPS = False
 except ImportError:
     USE_XFORMERS_OPS = False
 
@@ -200,7 +204,8 @@ class InternSdpaAttention(nn.Module):
         v = v.transpose(1, 2)
 
         x = F.scaled_dot_product_attention(q, k, v, scale=self.scale)
-        x = x.transpose(1, 2).view(B, N, -1)
+        # x = x.transpose(1, 2).view(B, N, -1)
+        x = x.transpose(1, 2).reshape(B, N, -1)
 
         x = self.proj(x)
         return x

vllm/model_executor/models/pixtral.py

@@ -8,8 +8,18 @@ import torch.nn.functional as F
 from mistral_common.protocol.instruct.messages import ImageChunk
 from PIL import Image
 from transformers import PretrainedConfig
-from xformers.ops.fmha import memory_efficient_attention
-from xformers.ops.fmha.attn_bias import BlockDiagonalMask
+# from xformers.ops.fmha import memory_efficient_attention
+# from xformers.ops.fmha.attn_bias import BlockDiagonalMask
+import vllm.envs as envs
+
+try:
+    if envs.VLLM_ATTENTION_BACKEND=="XFormers":
+        from xformers.ops.fmha import memory_efficient_attention
+        from xformers.ops.fmha.attn_bias import BlockDiagonalMask
+    else:
+        print("INFO: VLLM_ATTENTION_BACKEND is not XFormers.\n") 
+except ImportError:
+    print("INFO: Please install xformers if you want to infer pixtral.\n") 
 
 from vllm.attention import AttentionMetadata
 from vllm.config import CacheConfig, MultiModalConfig

vllm/model_executor/models/siglip.py

@@ -25,10 +25,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import SequenceData
+import vllm.envs as envs
 
 try:
-    from xformers import ops as xops
-    USE_XFORMERS_OPS = True
+    if envs.VLLM_ATTENTION_BACKEND=="XFormers":
+        from xformers import ops as xops
+        USE_XFORMERS_OPS = True
+    else:
+        USE_XFORMERS_OPS = False
 except ImportError:
     USE_XFORMERS_OPS = False