Fix numpy/meta ops for DPT/Dinov3 tests (#2325)

Author: lvyufeng

src/mindnlp/inference/layers/activation.py

@@ -4,10 +4,6 @@
 
 
 class SiluAndMul(nn.Module):
-
-    def __init__(self):
-        super().__init__()  # pylint: disable=useless-parent-delegation
-
     @mindtorch.compile
     def forward(self, x: mindtorch.Tensor) -> mindtorch.Tensor:
         x, y = x.chunk(2, -1)

src/mindnlp/inference/layers/attention.py

@@ -37,18 +37,18 @@ def forward(self, q: mindtorch.Tensor, k: mindtorch.Tensor, v: mindtorch.Tensor)
         k_cache, v_cache = self.k_cache, self.v_cache
         if k_cache.numel() and v_cache.numel():
             store_kvcache(k, v, k_cache, v_cache, context.slot_mapping)
-        if context.is_prefill:
-            if context.block_tables is not None:    # prefix cache
-                k, v = k_cache, v_cache
-            # pylint: disable=undefined-variable
-            o = flash_attn_varlen_func(q, k, v,
-                                       max_seqlen_q=context.max_seqlen_q, cu_seqlens_q=context.cu_seqlens_q,
-                                       max_seqlen_k=context.max_seqlen_k, cu_seqlens_k=context.cu_seqlens_k,
-                                       softmax_scale=self.scale, causal=True, block_table=context.block_tables)
-        else:    # decode
-            # flash_attn_with_kvcache is conditionally imported from flash_attn
-            # pylint: disable=undefined-variable
-            o = flash_attn_with_kvcache(q.unsqueeze(1), k_cache, v_cache,  # noqa: F821
-                                        cache_seqlens=context.context_lens, block_table=context.block_tables,
-                                        softmax_scale=self.scale, causal=True)
-        return o
+        # if context.is_prefill:
+        #     if context.block_tables is not None:    # prefix cache
+        #         k, v = k_cache, v_cache
+        #     # pylint: disable=undefined-variable
+        #     o = flash_attn_varlen_func(q, k, v,
+        #                                max_seqlen_q=context.max_seqlen_q, cu_seqlens_q=context.cu_seqlens_q,
+        #                                max_seqlen_k=context.max_seqlen_k, cu_seqlens_k=context.cu_seqlens_k,
+        #                                softmax_scale=self.scale, causal=True, block_table=context.block_tables)
+        # else:    # decode
+        #     # flash_attn_with_kvcache is conditionally imported from flash_attn
+        #     # pylint: disable=undefined-variable
+        #     o = flash_attn_with_kvcache(q.unsqueeze(1), k_cache, v_cache,  # noqa: F821
+        #                                 cache_seqlens=context.context_lens, block_table=context.block_tables,
+        #                                 softmax_scale=self.scale, causal=True)
+        # return o

src/mindnlp/inference/layers/sampler.py

@@ -3,10 +3,6 @@
 
 
 class Sampler(nn.Module):
-
-    def __init__(self):
-        super().__init__()  # pylint: disable=useless-parent-delegation
-
     def forward(self, logits: mindtorch.Tensor, temperatures: mindtorch.Tensor):
         logits = logits.to(mindtorch.float)
         greedy_tokens = logits.argmax(dim=-1)

src/mindnlp/patch/diffusers/__init__.py

@@ -23,4 +23,3 @@ def setup_diffusers_module():
     # Redirect mindnlp.diffusers to diffusers
     if 'mindnlp.diffusers' not in sys.modules:
         sys.modules['mindnlp.diffusers'] = diffusers
-

src/mindnlp/patch/diffusers/common.py

@@ -23,4 +23,3 @@ def patch_diffusers_common():
         )
     except ImportError:
         pass
-

src/mindnlp/patch/registry.py

@@ -237,4 +237,3 @@ def apply_all_patches(verbose: bool = False):
     apply_safetensors_patches(verbose=verbose)
     apply_transformers_patches(verbose=verbose)
     apply_diffusers_patches(verbose=verbose)
-

src/mindnlp/patch/transformers/__init__.py

@@ -38,4 +38,3 @@ def setup_transformers_module():
     transformers_module_name_nlp = 'mindnlp.transformers'
     if transformers_module_name_nlp not in sys.modules or not isinstance(sys.modules[transformers_module_name_nlp], _LazyModule):
         sys.modules[transformers_module_name_nlp] = lazy_module
-

src/mindnlp/patch/transformers/common.py

@@ -74,4 +74,3 @@ def empty_fn(*args, **kwargs):
     # Patch cache utils
     transformers.cache_utils.DynamicLayer.update = dynamic_layer_update
     transformers.cache_utils.DynamicSlidingWindowLayer.update = dynamic_sliding_window_layer_update
-

src/mindnlp/patch/transformers/v4_55.py

@@ -17,4 +17,3 @@ def patch_pre_trained_model_v4_55():
         'from_pretrained',
         [transformers.modeling_utils.restore_default_torch_dtype]
     )
-

src/mindnlp/patch/transformers/v4_56.py

@@ -17,4 +17,3 @@ def patch_pre_trained_model_v4_56():
         'from_pretrained',
         [transformers.modeling_utils.restore_default_dtype]
     )
-