remove input_ids from ForwardMeta (#4793)

fastdeploy/model_executor/forward_meta.py

@@ -65,8 +65,6 @@ class ForwardMeta:
     ForwardMeta is used to store the global meta information of the model forward.
     """
 
-    # Input tokens IDs
-    input_ids: paddle.Tensor
     # Input tokens IDs of removed padding
     ids_remove_padding: paddle.Tensor
     # Rotation position embedding
@@ -265,7 +263,7 @@ class HPUForwardMeta(ForwardMeta):
     """
 
     #
-    input_ids: paddle.Tensor
+    input_ids: paddle.Tensor = None
 
     # attention meta
     forward_mode: ForwardMode = ForwardMode.MIXED

fastdeploy/spec_decode/mtp.py

@@ -623,7 +623,6 @@ class MTPProposer(Proposer):
         """
         # Initialize forward meta
         self.forward_meta = ForwardMeta(
-            input_ids=self.model_inputs["input_ids"],
             ids_remove_padding=self.model_inputs["ids_remove_padding"],
             rotary_embs=self.model_inputs["rope_emb"],
             attn_backend=self.attn_backends[0],

fastdeploy/worker/gcu_model_runner.py

@@ -587,7 +587,6 @@ class GCUModelRunner(ModelRunnerBase):
         """
         # Initialize forward meta
         self.forward_meta = ForwardMeta(
-            input_ids=self.share_inputs["input_ids"],
             ids_remove_padding=self.share_inputs["ids_remove_padding"],
             rotary_embs=self.share_inputs["rope_emb"],
             attn_backend=self.attn_backends[0],

fastdeploy/worker/gpu_model_runner.py

@@ -1330,7 +1330,6 @@ class GPUModelRunner(ModelRunnerBase):
         """
         # Initialize forward meta
         self.forward_meta = ForwardMeta(
-            input_ids=self.share_inputs["input_ids"],
             ids_remove_padding=self.share_inputs["ids_remove_padding"],
             rotary_embs=self.share_inputs["rope_emb"],
             attn_backend=self.attn_backends[0],

fastdeploy/worker/metax_model_runner.py

@@ -1128,7 +1128,6 @@ class MetaxModelRunner(ModelRunnerBase):
         """
         # Initialize forward meta
         self.forward_meta = ForwardMeta(
-            input_ids=self.share_inputs["input_ids"],
             ids_remove_padding=self.share_inputs["ids_remove_padding"],
             rotary_embs=self.share_inputs["rope_emb"],
             attn_backend=self.attn_backends[0],

fastdeploy/worker/xpu_model_runner.py

@@ -89,7 +89,6 @@ def xpu_pre_process(
     share_inputs["cu_seqlens_k"] = cu_seqlens_k
 
     xpu_forward_meta = XPUForwardMeta(
-        input_ids=share_inputs["input_ids"],
         ids_remove_padding=share_inputs["ids_remove_padding"],
         rotary_embs=share_inputs["rope_emb"],
         attn_backend=None,

tests/graph_optimization/test_cuda_graph_dynamic_subgraph.py

@@ -42,12 +42,12 @@ class TestCase1SubLayer1(paddle.nn.Layer):
     def forward(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 forward pass"""
 
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
     def forward_correct(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 Correct forward pass"""
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
 
@@ -59,15 +59,15 @@ class TestCase1SubLayer2(paddle.nn.Layer):
 
     def forward(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer2 forward pass"""
-        x = forward_meta.input_ids
-        y = forward_meta.input_ids
+        x = forward_meta.ids_remove_padding
+        y = forward_meta.ids_remove_padding
         output = x + y
         return output
 
     def forward_correct(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer2 Correct forward pass"""
-        x = forward_meta.input_ids
-        y = forward_meta.input_ids
+        x = forward_meta.ids_remove_padding
+        y = forward_meta.ids_remove_padding
         output = x + y
         return output
 
@@ -81,12 +81,12 @@ class TestCase1SubLayer3(paddle.nn.Layer):
 
     def forward(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer3 forward pass"""
-        output = paddle.matmul(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.matmul(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
     def forward_correct(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer3 Correct forward pass"""
-        output = paddle.matmul(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.matmul(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
 
@@ -110,13 +110,12 @@ class TestModel1(paddle.nn.Layer):
         sublayer1_output = self.sublayer1(ids_remove_padding=ids_remove_padding, forward_meta=sub_meta1)
 
         # sublayer2 not use cuda graph
-        sub_meta2 = ForwardMeta(input_ids=sublayer1_output, ids_remove_padding=sublayer1_output)
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output)
         sublayer2_output = self.sublayer2(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
         self.sublayer2_output_buffer.copy_(sublayer2_output, False)
 
         # sublayer3 use cuda graph
         sub_meta3 = ForwardMeta(
-            input_ids=self.sublayer2_output_buffer,
             ids_remove_padding=self.sublayer2_output_buffer,
             step_use_cudagraph=True,
         )
@@ -134,11 +133,11 @@ class TestModel1(paddle.nn.Layer):
         )
 
         # sublayer2 not use cuda graph
-        sub_meta2 = ForwardMeta(input_ids=sublayer1_output, ids_remove_padding=sublayer1_output)
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output)
         sublayer2_output = self.sublayer2.forward_correct(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
 
         # sublayer3 not use cuda graph
-        sub_meta3 = ForwardMeta(input_ids=sublayer2_output, ids_remove_padding=sublayer2_output)
+        sub_meta3 = ForwardMeta(ids_remove_padding=sublayer2_output)
         sublayer3_output = self.sublayer3.forward_correct(ids_remove_padding=sublayer2_output, forward_meta=sub_meta3)
 
         return sublayer3_output
@@ -175,7 +174,7 @@ class TestCUDAGrpahSubgraph(unittest.TestCase):
         # Run Test Case1
         test_model1 = TestModel1(fd_config=fd_config)
         input_tensor1 = paddle.ones([8])
-        forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
+        forward_meta1 = ForwardMeta(ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
         # Trigger Capture
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)

tests/graph_optimization/test_cuda_graph_recapture.py

@@ -27,13 +27,13 @@ class TestCase1SubLayer1(paddle.nn.Layer):
     def forward(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 forward pass"""
 
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
     def forward_correct(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 Correct forward pass"""
 
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
 
@@ -55,9 +55,7 @@ class TestModel1(paddle.nn.Layer):
         sublayer1_output = self.sublayer1(ids_remove_padding=ids_remove_padding, forward_meta=sub_meta1)
 
         # sublayer2 use cuda graph
-        sub_meta2 = ForwardMeta(
-            input_ids=sublayer1_output, ids_remove_padding=sublayer1_output, step_use_cudagraph=True
-        )
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output, step_use_cudagraph=True)
         sublayer2_output = self.sublayer2(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
 
         return sublayer2_output
@@ -71,7 +69,7 @@ class TestModel1(paddle.nn.Layer):
         )
 
         # sublayer2 not use cuda graph
-        sub_meta2 = ForwardMeta(input_ids=sublayer1_output, ids_remove_padding=sublayer1_output)
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output)
         sublayer2_output = self.sublayer2.forward_correct(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
 
         return sublayer2_output
@@ -109,7 +107,7 @@ class TestCUDAGrpahRecapture(unittest.TestCase):
         # Run Test Case1
         self.test_model1 = TestModel1(fd_config=fd_config)
         input_tensor1 = paddle.ones([1, 32768])
-        forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
+        forward_meta1 = ForwardMeta(ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
         # Correct output
         self.output_correct = self.test_model1.forward_correct(

tests/graph_optimization/test_cuda_graph_spec_decode.py

@@ -42,13 +42,13 @@ class TestCase1SubLayer1(paddle.nn.Layer):
     def forward(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 forward pass"""
 
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
     def forward_correct(self, ids_remove_padding, forward_meta: ForwardMeta):
         """Sub layer1 Correct forward pass"""
 
-        output = paddle.add(forward_meta.input_ids, forward_meta.input_ids)
+        output = paddle.add(forward_meta.ids_remove_padding, forward_meta.ids_remove_padding)
         return output
 
 
@@ -70,9 +70,7 @@ class TestModel1(paddle.nn.Layer):
         sublayer1_output = self.sublayer1(ids_remove_padding=ids_remove_padding, forward_meta=sub_meta1)
 
         # sublayer2 use cuda graph
-        sub_meta2 = ForwardMeta(
-            input_ids=sublayer1_output, ids_remove_padding=sublayer1_output, step_use_cudagraph=True
-        )
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output, step_use_cudagraph=True)
         sublayer2_output = self.sublayer2(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
 
         return sublayer2_output
@@ -86,7 +84,7 @@ class TestModel1(paddle.nn.Layer):
         )
 
         # sublayer2 not use cuda graph
-        sub_meta2 = ForwardMeta(input_ids=sublayer1_output, ids_remove_padding=sublayer1_output)
+        sub_meta2 = ForwardMeta(ids_remove_padding=sublayer1_output)
         sublayer2_output = self.sublayer2.forward_correct(ids_remove_padding=sublayer1_output, forward_meta=sub_meta2)
 
         return sublayer2_output
@@ -122,7 +120,7 @@ class TestCUDAGrpahSpecDecode(unittest.TestCase):
         # Run Test Case1
         test_model1 = TestModel1(fd_config=fd_config)
         input_tensor1 = paddle.ones([1, 32768])
-        forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
+        forward_meta1 = ForwardMeta(ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
         # Trigger Capture
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)

tests/graph_optimization/test_graph_opt_backend.py

@@ -108,9 +108,7 @@ class TestGraphOptBackend(unittest.TestCase):
 
         # Create input data
         self.input_tensor = paddle.randint(32, shape=self.input_shape, dtype=self.dtype)
-        self.forward_meta = ForwardMeta(
-            input_ids=self.input_tensor, ids_remove_padding=self.input_tensor, step_use_cudagraph=True
-        )
+        self.forward_meta = ForwardMeta(ids_remove_padding=self.input_tensor, step_use_cudagraph=True)
 
         # Compute baseline result once
         baseline_model = Attention(fd_config=self.baseline_fd_config, **self.model_config)

tests/graph_optimization/test_static_graph_cuda_graph_split.py

@@ -108,7 +108,7 @@ class TestStaticGraphCUDAGraphSplit(unittest.TestCase):
 
         test_model1 = TestModel(fd_config=fd_config)
         x = paddle.randint(32, shape=[1, 8])
-        forward_meta1 = ForwardMeta(input_ids=x, ids_remove_padding=x, step_use_cudagraph=True)
+        forward_meta1 = ForwardMeta(ids_remove_padding=x, step_use_cudagraph=True)
 
         # Trigger Capture
         with sot_warmup_guard(True):

tests/layers/test_attention_layer.py

@@ -320,7 +320,6 @@ class TestAttentionPerformance(unittest.TestCase):
         ) = pre_process(input_ids, seq_lens_this_time, False, None, seq_lens_encoder, seq_lens_decoder)
 
         meta = ForwardMeta(
-            input_ids=input_ids,
             ids_remove_padding=ids_remove_padding,
             seq_lens_encoder=seq_lens_encoder,
             seq_lens_decoder=seq_lens_decoder,

tests/model_executor/test_forward_meta_str.py

@@ -30,6 +30,7 @@ class TOYGPUModelRunner:
         self.pre_max_block_num = 16
         # Not the tensor in real sense, just for make ForwardMeta
         self.share_inputs = {}
+
         self.share_inputs["input_ids"] = paddle.full(
             [self.max_num_seqs, self.max_model_len],
             0,
@@ -63,7 +64,6 @@ class TOYGPUModelRunner:
         """
         # Ignore the attentionbackbend for simplify
         self.forward_meta = ForwardMeta(
-            input_ids=self.share_inputs["input_ids"],
             ids_remove_padding=self.share_inputs["ids_remove_padding"],
             # rotary_embs=self.share_inputs["rope_emb"],# Ignore the rope_emb for simplify
             # attn_backend=self.attn_backends[0],# Ignore the attn_backbend for simplify