Compilation Folder Reset (#11539)

Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>

python/sglang/srt/model_executor/compilation/backend.py → python/sglang/srt/compilation/backend.py

@@ -15,15 +15,11 @@ import torch
 import torch.fx as fx
 from torch._dispatch.python import enable_python_dispatcher
 
-from sglang.srt.model_executor.compilation.compilation_config import CompilationConfig
-from sglang.srt.model_executor.compilation.compilation_counter import (
-    compilation_counter,
-)
-from sglang.srt.model_executor.compilation.compiler_interface import InductorAdaptor
-from sglang.srt.model_executor.compilation.cuda_piecewise_backend import (
-    CUDAPiecewiseBackend,
-)
-from sglang.srt.model_executor.compilation.pass_manager import PostGradPassManager
+from sglang.srt.compilation.compilation_config import CompilationConfig
+from sglang.srt.compilation.compilation_counter import compilation_counter
+from sglang.srt.compilation.compiler_interface import InductorAdaptor
+from sglang.srt.compilation.cuda_piecewise_backend import CUDAPiecewiseBackend
+from sglang.srt.compilation.pass_manager import PostGradPassManager
 
 logger = logging.getLogger(__name__)
 

python/sglang/srt/model_executor/compilation/compile.py → python/sglang/srt/compilation/compile.py

@@ -10,7 +10,7 @@ from typing import Any, Callable, Optional, Union
 
 import torch
 
-from sglang.srt.model_executor.compilation.compilation_config import CompilationConfig
+from sglang.srt.compilation.compilation_config import CompilationConfig
 
 logger = logging.getLogger(__name__)
 
@@ -134,7 +134,7 @@ def install_torch_compiled(
     dyn_map = dynamic_arg_dims or _infer_dynamic_arg_dims_from_annotations(unbound_fwd)
 
     if backend_factory is None:
-        from sglang.srt.model_executor.compilation.backend import SGLangBackend
+        from sglang.srt.compilation.backend import SGLangBackend
 
         backend_factory = lambda gm, ex: SGLangBackend(compile_config, graph_pool)(
             gm, ex

python/sglang/srt/model_executor/compilation/compiler_interface.py → python/sglang/srt/compilation/compiler_interface.py

@@ -12,10 +12,8 @@ import torch
 import torch._inductor.compile_fx
 import torch.fx as fx
 
-from sglang.srt.model_executor.compilation.compilation_counter import (
-    compilation_counter,
-)
-from sglang.srt.model_executor.compilation.inductor_pass import pass_context
+from sglang.srt.compilation.compilation_counter import compilation_counter
+from sglang.srt.compilation.inductor_pass import pass_context
 
 
 class CompilerInterface:

python/sglang/srt/model_executor/compilation/cuda_piecewise_backend.py → python/sglang/srt/compilation/cuda_piecewise_backend.py

@@ -9,11 +9,9 @@ from unittest.mock import patch
 import torch
 import torch.fx as fx
 
-import sglang.srt.model_executor.compilation.weak_ref_tensor_jit
-from sglang.srt.model_executor.compilation.compilation_config import CompilationConfig
-from sglang.srt.model_executor.compilation.compilation_counter import (
-    compilation_counter,
-)
+import sglang.srt.compilation.weak_ref_tensor_jit
+from sglang.srt.compilation.compilation_config import CompilationConfig
+from sglang.srt.compilation.compilation_counter import compilation_counter
 
 logger = logging.getLogger(__name__)
 

python/sglang/srt/model_executor/compilation/fix_functionalization.py → python/sglang/srt/compilation/fix_functionalization.py

@@ -8,8 +8,8 @@ from typing import Optional, Union
 import torch
 from torch._higher_order_ops.auto_functionalize import auto_functionalized
 
-from sglang.srt.model_executor.compilation.fx_utils import is_func
-from sglang.srt.model_executor.compilation.inductor_pass import SGLangInductorPass
+from sglang.srt.compilation.fx_utils import is_func
+from sglang.srt.compilation.inductor_pass import SGLangInductorPass
 
 logger = logging.getLogger(__name__)
 

python/sglang/srt/model_executor/compilation/pass_manager.py → python/sglang/srt/compilation/pass_manager.py

@@ -4,10 +4,8 @@ import logging
 
 from torch import fx as fx
 
-from sglang.srt.model_executor.compilation.fix_functionalization import (
-    FixFunctionalizationPass,
-)
-from sglang.srt.model_executor.compilation.inductor_pass import (
+from sglang.srt.compilation.fix_functionalization import FixFunctionalizationPass
+from sglang.srt.compilation.inductor_pass import (
     CustomGraphPass,
     InductorPass,
     SGLangInductorPass,

python/sglang/srt/distributed/parallel_state.py

@@ -337,7 +337,6 @@ class GroupCoordinator:
             else:
                 ca_max_size = 8 * 1024 * 1024
             try:
-                # print(f"ca_max_size: {ca_max_size}")
                 self.ca_comm = CustomAllreduce(
                     group=self.cpu_group,
                     device=self.device,

python/sglang/srt/layers/radix_attention.py

@@ -20,15 +20,13 @@ from typing import TYPE_CHECKING, Optional
 import torch
 from torch import nn
 
+from sglang.srt.compilation.piecewise_context_manager import get_forward_context
+from sglang.srt.utils import direct_register_custom_op
+
 if TYPE_CHECKING:
     from sglang.srt.layers.quantization.base_config import QuantizationConfig
     from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
-from sglang.srt.model_executor.compilation.piecewise_context_manager import (
-    get_forward_context,
-)
-from sglang.srt.utils import direct_register_custom_op
-
 
 class AttentionType(Enum):
     """
@@ -112,7 +110,7 @@ class RadixAttention(nn.Module):
                 k = k.view(-1, self.tp_k_head_num, self.v_head_dim)
 
         if forward_batch.forward_mode.is_extend() and get_forward_context() is not None:
-            output = torch.zeros_like(q)
+            output = torch.empty_like(q)
             torch.ops.sglang.unified_attention_with_output(
                 q, k, v, output, save_kv_cache, self.layer_id
             )

python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py

@@ -24,6 +24,9 @@ from typing import TYPE_CHECKING, Union
 import torch
 import tqdm
 
+from sglang.srt.compilation.compilation_config import CompilationConfig
+from sglang.srt.compilation.compile import install_torch_compiled, set_compiled
+from sglang.srt.compilation.piecewise_context_manager import set_forward_context
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.distributed import get_tensor_model_parallel_rank
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
@@ -38,14 +41,6 @@ from sglang.srt.layers.dp_attention import (
 )
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.layers.torchao_utils import save_gemlite_cache
-from sglang.srt.model_executor.compilation.compilation_config import CompilationConfig
-from sglang.srt.model_executor.compilation.compile import (
-    install_torch_compiled,
-    set_compiled,
-)
-from sglang.srt.model_executor.compilation.piecewise_context_manager import (
-    set_forward_context,
-)
 from sglang.srt.model_executor.forward_batch_info import (
     CaptureHiddenMode,
     ForwardBatch,