sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/compilation/cuda_piecewise_backend.py

@@ -0,0 +1,228 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.10.0/vllm/compilation/cuda_piecewise_backend.py
+
+import dataclasses
+import logging
+from contextlib import ExitStack
+from typing import Any, Callable, Optional, Union
+from unittest.mock import patch
+
+import torch
+import torch.fx as fx
+
+import sglang.srt.compilation.weak_ref_tensor_jit  # noqa: F401
+from sglang.srt.compilation.compilation_config import CompilationConfig
+from sglang.srt.compilation.compilation_counter import compilation_counter
+
+logger = logging.getLogger(__name__)
+
+
+def weak_ref_tensor(tensor: Any) -> Any:
+    """
+    Create a weak reference to a tensor.
+    The new tensor will share the same data as the original tensor,
+    but will not keep the original tensor alive.
+    """
+    if isinstance(tensor, torch.Tensor):
+        # TODO(yuwei): introduce weak_ref_tensor from sgl_kernel
+        return torch.ops.jit_weak_ref_tensor.weak_ref_tensor(tensor)
+    return tensor
+
+
+def weak_ref_tensors(
+    tensors: Union[torch.Tensor, list[torch.Tensor], tuple[torch.Tensor]]
+) -> Union[torch.Tensor, list[Any], tuple[Any], Any]:
+    """
+    Convenience function to create weak references to tensors,
+    for single tensor, list of tensors or tuple of tensors.
+    """
+    if isinstance(tensors, torch.Tensor):
+        return weak_ref_tensor(tensors)
+    if isinstance(tensors, list):
+        return [weak_ref_tensor(t) for t in tensors]
+    if isinstance(tensors, tuple):
+        return tuple(weak_ref_tensor(t) for t in tensors)
+    raise ValueError("Invalid type for tensors")
+
+
+@dataclasses.dataclass
+class ConcreteSizeEntry:
+    runtime_shape: int
+    need_to_compile: bool  # the size is in compile_sizes
+    use_cudagraph: bool  # the size is in cudagraph_capture_sizes
+
+    compiled: bool = False
+    runnable: Callable = None  # type: ignore
+    num_finished_warmup: int = 0
+    cudagraph: Optional[torch.cuda.CUDAGraph] = None
+    output: Optional[Any] = None
+
+    # for cudagraph debugging, track the input addresses
+    # during capture, and check if they are the same during replay
+    input_addresses: Optional[list[int]] = None
+
+
+class CUDAPiecewiseBackend:
+
+    def __init__(
+        self,
+        graph: fx.GraphModule,
+        compile_config: CompilationConfig,
+        inductor_config: dict[str, Any],
+        graph_pool: Any,
+        piecewise_compile_index: int,
+        total_piecewise_compiles: int,
+        sym_shape_indices: list[int],
+        compiled_graph_for_general_shape: Callable,
+        sglang_backend,
+    ):
+        """
+        The backend for piecewise compilation.
+        It mainly handles the compilation and cudagraph capturing.
+
+        We will compile `self.graph` once for the general shape,
+        and then compile for different shapes specified in
+        `compilation_config.compile_sizes`.
+
+        Independently, we will capture cudagraph for different shapes.
+
+        If a shape needs both compilation and cudagraph, we will
+        compile it first, and then capture cudagraph.
+        """
+        self.graph = graph
+        self.inductor_config = inductor_config
+        self.graph_pool = graph_pool
+        self.piecewise_compile_index = piecewise_compile_index
+        self.total_piecewise_compiles = total_piecewise_compiles
+        self.sglang_backend = sglang_backend
+
+        self.is_first_graph = piecewise_compile_index == 0
+        self.is_last_graph = piecewise_compile_index == total_piecewise_compiles - 1
+
+        self.compile_sizes: set[int] = set([])
+        self.compile_config = compile_config
+        self.cudagraph_capture_sizes: set[int] = set(compile_config.get_capture_sizes())
+
+        self.first_run_finished = False
+
+        self.compiled_graph_for_general_shape = compiled_graph_for_general_shape  # noqa
+
+        self.sym_shape_indices = sym_shape_indices
+
+        self.is_debugging_mode = True
+
+        # the entries for different shapes that we need to either
+        # compile or capture cudagraph
+        self.concrete_size_entries: dict[int, ConcreteSizeEntry] = {}
+
+        # to_be_compiled_sizes tracks the remaining sizes to compile,
+        # and updates during the compilation process, so we need to copy it
+        self.to_be_compiled_sizes: set[int] = self.compile_sizes.copy()
+        for shape in self.compile_sizes.union(self.cudagraph_capture_sizes):
+            self.concrete_size_entries[shape] = ConcreteSizeEntry(
+                runtime_shape=shape,
+                need_to_compile=shape in self.compile_sizes,
+                use_cudagraph=shape in self.cudagraph_capture_sizes,
+            )
+
+    def check_for_ending_compilation(self):
+        if self.is_last_graph and not self.to_be_compiled_sizes:
+            # no specific sizes to compile
+            # save the hash of the inductor graph for the next run
+            self.sglang_backend.compiler_manager.save_to_file()
+
+    def __call__(self, *args) -> Any:
+        if not self.first_run_finished:
+            self.first_run_finished = True
+            self.check_for_ending_compilation()
+            return self.compiled_graph_for_general_shape(*args)
+        runtime_shape = args[self.sym_shape_indices[0]]
+        if runtime_shape not in self.concrete_size_entries:
+            # we don't need to do anything for this shape
+            return self.compiled_graph_for_general_shape(*args)
+
+        entry = self.concrete_size_entries[runtime_shape]
+
+        if entry.runnable is None:
+            entry.runnable = self.compiled_graph_for_general_shape
+
+        if entry.need_to_compile and not entry.compiled:
+            entry.compiled = True
+            self.to_be_compiled_sizes.remove(runtime_shape)
+            # args are real arguments
+            entry.runnable = self.sglang_backend.compiler_manager.compile(
+                self.graph,
+                args,
+                self.inductor_config,
+                graph_index=self.piecewise_compile_index,
+                num_graphs=self.total_piecewise_compiles,
+                runtime_shape=runtime_shape,
+            )
+
+            # finished compilations for all required shapes
+            if self.is_last_graph and not self.to_be_compiled_sizes:
+                self.check_for_ending_compilation()
+
+        # Skip CUDA graphs if this entry doesn't use them OR
+        # if we're supposed to skip them globally
+        # skip_cuda_graphs = get_forward_context().skip_cuda_graphs
+        # if not entry.use_cudagraph or skip_cuda_graphs:
+        #     return entry.runnable(*args)
+
+        if entry.cudagraph is None:
+            if entry.num_finished_warmup < 1:  # noqa
+                entry.num_finished_warmup += 1
+                return entry.runnable(*args)
+
+            input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            entry.input_addresses = input_addresses
+            cudagraph = torch.cuda.CUDAGraph()
+
+            with ExitStack() as stack:
+                if not self.is_first_graph:
+                    # during every model forward, we will capture
+                    # many pieces of cudagraphs (roughly one per layer).
+                    # running gc again and again across layers will
+                    # make the cudagraph capture very slow.
+                    # therefore, we only run gc for the first graph,
+                    # and disable gc for the rest of the graphs.
+                    stack.enter_context(patch("gc.collect", lambda: None))
+                    stack.enter_context(patch("torch.cuda.empty_cache", lambda: None))
+
+                # mind-exploding: carefully manage the reference and memory.
+                with torch.cuda.graph(cudagraph, pool=self.graph_pool):
+                    # `output` is managed by pytorch's cudagraph pool
+                    output = entry.runnable(*args)
+                    if self.is_last_graph:
+                        # by converting it to weak ref,
+                        # the original `output` will immediately be released
+                        # to save memory. It is only safe to do this for
+                        # the last graph, because the output of the last graph
+                        # will not be used by any other cuda graph.
+                        output = weak_ref_tensors(output)
+
+            # here we always use weak ref for the output
+            # to save memory
+            entry.output = weak_ref_tensors(output)
+            entry.cudagraph = cudagraph
+
+            compilation_counter.num_cudagraph_captured += 1
+
+            # important: we need to return the output, rather than
+            # the weak ref of the output, so that pytorch can correctly
+            # manage the memory during cuda graph capture
+            return output
+
+        if self.is_debugging_mode:
+            # check if the input addresses are the same
+            new_input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            assert new_input_addresses == entry.input_addresses, (
+                "Input addresses for cudagraphs are different during replay."
+                f" Expected {entry.input_addresses}, got {new_input_addresses}"
+            )
+
+        entry.cudagraph.replay()
+        return entry.output

sglang/srt/compilation/fix_functionalization.py

@@ -0,0 +1,134 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.10.0/vllm/compilation/fix_functionalization.py
+
+import logging
+import operator
+from collections.abc import Iterable
+from typing import Optional, Union
+
+import torch
+from torch._higher_order_ops.auto_functionalize import auto_functionalized
+
+from sglang.srt.compilation.fx_utils import is_func
+from sglang.srt.compilation.inductor_pass import SGLangInductorPass
+
+logger = logging.getLogger(__name__)
+
+
+class FixFunctionalizationPass(SGLangInductorPass):
+    """
+    This pass defunctionalizes certain nodes to avoid redundant tensor copies.
+    After this pass, DCE (dead-code elimination) should never be run,
+    as de-functionalized nodes may appear as dead code.
+
+    To add new nodes to defunctionalize, add to the if-elif chain in __call__.
+    """
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.begin()
+        self.dump_graph(graph, "before_fix_functionalization")
+
+        self.nodes_to_remove: list[torch.fx.Node] = []
+        count = 0
+        for node in graph.nodes:
+            if not is_func(node, auto_functionalized):
+                continue  # Avoid deep if-elif nesting
+            count += 1
+
+        self.dump_graph(graph, "before_fix_functionalization_cleanup")
+
+        # Remove the nodes all at once
+        count_removed = len(self.nodes_to_remove)
+        for node in self.nodes_to_remove:
+            graph.erase_node(node)
+
+        logger.debug(
+            "De-functionalized %s nodes, removed %s nodes", count, count_removed
+        )
+        self.dump_graph(graph, "after_fix_functionalization")
+        self.end_and_log()
+
+    def _remove(self, node_or_nodes: Union[torch.fx.Node, Iterable[torch.fx.Node]]):
+        """
+        Stage a node (or nodes) for removal at the end of the pass.
+        """
+        if isinstance(node_or_nodes, torch.fx.Node):
+            self.nodes_to_remove.append(node_or_nodes)
+        else:
+            self.nodes_to_remove.extend(node_or_nodes)
+
+    def defunctionalize(
+        self,
+        graph: torch.fx.Graph,
+        node: torch.fx.Node,
+        mutated_args: dict[int, Union[torch.fx.Node, str]],
+        args: Optional[tuple[Union[torch.fx.Node, str], ...]] = None,
+    ):
+        """
+        De-functionalize a node by replacing it with a call to the original.
+        It also replaces the getitem users with the mutated arguments.
+        See replace_users_with_mutated_args and insert_defunctionalized.
+        """
+        self.replace_users_with_mutated_args(node, mutated_args)
+        self.insert_defunctionalized(graph, node, args=args)
+        self._remove(node)
+
+    def replace_users_with_mutated_args(
+        self, node: torch.fx.Node, mutated_args: dict[int, Union[torch.fx.Node, str]]
+    ):
+        """
+        Replace all getitem users of the auto-functionalized node with the
+        mutated arguments.
+        :param node: The auto-functionalized node
+        :param mutated_args: The mutated arguments, indexed by getitem index.
+        If the value of an arg is a string, `node.kwargs[arg]` is used.
+        """
+        for idx, user in self.getitem_users(node).items():
+            arg = mutated_args[idx]
+            arg = node.kwargs[arg] if isinstance(arg, str) else arg
+            user.replace_all_uses_with(arg)
+            self._remove(user)
+
+    def getitem_users(self, node: torch.fx.Node) -> dict[int, torch.fx.Node]:
+        """
+        Returns the operator.getitem users of the auto-functionalized node,
+        indexed by the index they are getting.
+        """
+        users = {}
+        for user in node.users:
+            if is_func(user, operator.getitem):
+                idx = user.args[1]
+                users[idx] = user
+        return users
+
+    def insert_defunctionalized(
+        self,
+        graph: torch.fx.Graph,
+        node: torch.fx.Node,
+        args: Optional[tuple[Union[torch.fx.Node, str], ...]] = None,
+    ):
+        """
+        Insert a new defunctionalized node into the graph before node.
+        If one of the kwargs is 'out', provide args directly,
+        as node.kwargs cannot be used.
+        See https://github.com/pytorch/pytorch/blob/a00faf440888ffb724bad413f329a49e2b6388e7/torch/_inductor/lowering.py#L351
+
+        :param graph: Graph to insert the defunctionalized node into
+        :param node: The auto-functionalized node to defunctionalize
+        :param args: If we cannot use kwargs, specify args directly.
+        If an arg is a string, `node.kwargs[arg]` is used.
+        """  # noqa: E501
+        assert is_func(
+            node, auto_functionalized
+        ), f"node must be auto-functionalized, is {node} instead"
+
+        # Create a new call to the original function
+        with graph.inserting_before(node):
+            function = node.args[0]
+            if args is None:
+                graph.call_function(function, kwargs=node.kwargs)
+            else:
+                # Args passed as strings refer to items in node.kwargs
+                args = tuple(
+                    node.kwargs[arg] if isinstance(arg, str) else arg for arg in args
+                )
+                graph.call_function(function, args=args)

sglang/srt/compilation/fx_utils.py

@@ -0,0 +1,83 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.10.0/vllm/compilation/fx_utils.py
+
+import operator
+from collections.abc import Iterable, Iterator
+from typing import Optional
+
+from torch import fx
+from torch._higher_order_ops.auto_functionalize import auto_functionalized
+from torch._ops import OpOverload
+
+
+def is_func(node: fx.Node, target) -> bool:
+    return node.op == "call_function" and node.target == target
+
+
+def is_auto_func(node: fx.Node, op: OpOverload) -> bool:
+    return is_func(node, auto_functionalized) and node.args[0] == op
+
+
+# Returns the first specified node with the given op (if it exists)
+def find_specified_fn_maybe(
+    nodes: Iterable[fx.Node], op: OpOverload
+) -> Optional[fx.Node]:
+    for node in nodes:
+        if node.target == op:
+            return node
+    return None
+
+
+# Returns the first specified node with the given op
+def find_specified_fn(nodes: Iterable[fx.Node], op: OpOverload) -> fx.Node:
+    node = find_specified_fn_maybe(nodes, op)
+    assert node is not None, f"Could not find {op} in nodes {nodes}"
+    return node
+
+
+# Returns the first auto_functionalized node with the given op (if it exists)
+def find_auto_fn_maybe(nodes: Iterable[fx.Node], op: OpOverload) -> Optional[fx.Node]:
+    for node in nodes:
+        if is_func(node, auto_functionalized) and node.args[0] == op:  # noqa
+            return node
+    return None
+
+
+# Returns the first auto_functionalized node with the given op
+def find_auto_fn(nodes: Iterable[fx.Node], op: OpOverload) -> fx.Node:
+    node = find_auto_fn_maybe(nodes, op)
+    assert node is not None, f"Could not find {op} in nodes {nodes}"
+    return node
+
+
+# Returns the getitem node that extracts the idx-th element from node
+# (if it exists)
+def find_getitem_maybe(node: fx.Node, idx: int) -> Optional[fx.Node]:
+    for user in node.users:
+        if is_func(user, operator.getitem) and user.args[1] == idx:
+            return user
+    return None
+
+
+# Returns the getitem node that extracts the idx-th element from node
+def find_getitem(node: fx.Node, idx: int) -> fx.Node:
+    ret = find_getitem_maybe(node, idx)
+    assert ret is not None, f"Could not find getitem {idx} in node {node}"
+    return ret
+
+
+# An auto-functionalization-aware utility for finding nodes with a specific op
+def find_op_nodes(op: OpOverload, graph: fx.Graph) -> Iterator[fx.Node]:
+    if not op._schema.is_mutable:
+        yield from graph.find_nodes(op="call_function", target=op)
+
+    for n in graph.find_nodes(op="call_function", target=auto_functionalized):
+        if n.args[0] == op:
+            yield n
+
+
+# Asserts that the node only has one user and returns it
+# Even if a node has only 1 user, it might share storage with another node,
+# which might need to be taken into account.
+def get_only_user(node: fx.Node) -> fx.Node:
+    assert len(node.users) == 1
+    return next(iter(node.users))

sglang/srt/compilation/inductor_pass.py

@@ -0,0 +1,140 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.10.0/vllm/compilation/inductor_pass.py
+
+import hashlib
+import inspect
+import json
+import logging
+import time
+import types
+from contextlib import contextmanager
+from typing import Any, Callable, Optional, Union
+
+import torch
+from torch import fx
+from torch._dynamo.utils import lazy_format_graph_code
+from torch._inductor.custom_graph_pass import CustomGraphPass
+
+logger = logging.getLogger(__name__)
+
+_pass_context = None
+
+
+class PassContext:
+
+    def __init__(self, runtime_shape: Optional[int]):
+        self.runtime_shape = runtime_shape
+
+
+def get_pass_context() -> PassContext:
+    """Get the current pass context."""
+    assert _pass_context is not None
+    return _pass_context
+
+
+@contextmanager
+def pass_context(runtime_shape: Optional[int]):
+    """A context manager that stores the current pass context,
+    usually it is a list of sizes to specialize.
+    """
+    global _pass_context
+    prev_context = _pass_context
+    _pass_context = PassContext(runtime_shape)
+    try:
+        yield
+    finally:
+        _pass_context = prev_context
+
+
+class InductorPass(CustomGraphPass):
+    """
+    A custom graph pass that uses a hash of its source as the UUID.
+    This is defined as a convenience and should work in most cases.
+    """
+
+    def uuid(self) -> Any:
+        """
+        Provide a unique identifier for the pass, used in Inductor code cache.
+        This should depend on the pass implementation, so that changes to the
+        pass result in recompilation.
+        By default, the object source is hashed.
+        """
+        return InductorPass.hash_source(self)
+
+    @staticmethod
+    def hash_source(*srcs: Union[str, Any]):
+        """
+        Utility method to hash the sources of functions or objects.
+        :param srcs: strings or objects to add to the hash.
+        Objects and functions have their source inspected.
+        :return:
+        """
+        hasher = hashlib.sha256()
+        for src in srcs:
+            if isinstance(src, str):
+                src_str = src
+            elif isinstance(src, types.FunctionType):
+                src_str = inspect.getsource(src)
+            else:
+                src_str = inspect.getsource(src.__class__)
+            hasher.update(src_str.encode("utf-8"))
+        return hasher.hexdigest()
+
+    @staticmethod
+    def hash_dict(dict_: dict[Any, Any]):
+        """
+        Utility method to hash a dictionary, can alternatively be used for uuid.
+        :return: A sha256 hash of the json rep of the dictionary.
+        """
+        encoded = json.dumps(dict_, sort_keys=True).encode("utf-8")
+        return hashlib.sha256(encoded).hexdigest()
+
+    def is_applicable_for_shape(self, shape: Optional[int]):
+        return True
+
+
+class CallableInductorPass(InductorPass):
+    """
+    This class is a wrapper for a callable that automatically provides an
+    implementation of the UUID.
+    """
+
+    def __init__(
+        self, callable: Callable[[fx.Graph], None], uuid: Optional[Any] = None
+    ):
+        self.callable = callable
+        self._uuid = self.hash_source(callable) if uuid is None else uuid
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.callable(graph)
+
+    def uuid(self) -> Any:
+        return self._uuid
+
+
+class SGLangInductorPass(InductorPass):
+
+    def __init__(
+        self,
+    ):
+        self.pass_name = self.__class__.__name__
+
+    def dump_graph(self, graph: torch.fx.Graph, stage: str):
+        lazy_format_graph_code(stage, graph.owning_module)
+
+    def begin(self):
+        self._start_time = time.perf_counter_ns()
+
+    def end_and_log(self):
+        self._end_time = time.perf_counter_ns()
+        duration_ms = float(self._end_time - self._start_time) / 1.0e6
+        logger.debug("%s completed in %.1f ms", self.pass_name, duration_ms)
+
+
+class PrinterInductorPass(SGLangInductorPass):
+
+    def __init__(self, name: str):
+        super().__init__()
+        self.name = name
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.dump_graph(graph, self.name)

sglang/srt/compilation/pass_manager.py

@@ -0,0 +1,66 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.10.0/vllm/compilation/pass_manager.py
+
+import logging
+
+from torch import fx as fx
+
+from sglang.srt.compilation.fix_functionalization import FixFunctionalizationPass
+from sglang.srt.compilation.inductor_pass import (
+    CustomGraphPass,
+    InductorPass,
+    SGLangInductorPass,
+    get_pass_context,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class PostGradPassManager(CustomGraphPass):
+    """
+    The pass manager for post-grad passes.
+    It handles configuration, adding custom passes, and running passes.
+    It supports uuid for the Inductor code cache. That includes torch<2.6
+    support using pickling (in .inductor_pass.CustomGraphPass).
+
+    The order of the post-grad post-passes is:
+    1. passes (constructor parameter)
+    2. default passes (NoopEliminationPass, FusionPass)
+    3. config["post_grad_custom_post_pass"] (if it exists)
+    4. fix_functionalization
+    This way, all passes operate on a functionalized graph.
+    """
+
+    def __init__(self):
+        self.passes: list[SGLangInductorPass] = []
+
+    def __call__(self, graph: fx.Graph):
+        shape = get_pass_context().runtime_shape
+        for pass_ in self.passes:
+            if pass_.is_applicable_for_shape(shape):
+                pass_(graph)
+
+        # always run fix_functionalization last
+        self.fix_functionalization(graph)
+
+    def configure(
+        self,
+    ):
+        self.pass_config = dict()
+        self.fix_functionalization = FixFunctionalizationPass()
+
+    def add(self, pass_: InductorPass):
+        assert isinstance(pass_, InductorPass)
+        self.passes.append(pass_)
+
+    def uuid(self):
+        """
+        The PostGradPassManager is set as a custom pass in the Inductor and
+        affects compilation caching. Its uuid depends on the UUIDs of all
+        dependent passes and the pass config. See InductorPass for more info.
+        """
+        pass_manager_uuid = "fshdakhsa"
+        state = {"pass_config": pass_manager_uuid, "passes": []}
+        for pass_ in self.passes:
+            state["passes"].append(pass_.uuid())
+        state["passes"].append(self.fix_functionalization.uuid())
+        return InductorPass.hash_dict(state)

sglang/srt/compilation/piecewise_context_manager.py

@@ -0,0 +1,40 @@
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+
+
+@dataclass
+class ForwardContext:
+    def __init__(self):
+        self.forward_batch = None
+        self.attention_layer = None
+
+    def set_forward_batch(self, forward_batch: ForwardBatch):
+        self.forward_batch = forward_batch
+
+    def set_attention_layers(self, layers: List[Any]):
+        self.attention_layers = layers
+
+
+_forward_context: Optional[ForwardContext] = None
+
+
+def get_forward_context() -> Optional[ForwardContext]:
+    if _forward_context is None:
+        return None
+    return _forward_context
+
+
+@contextmanager
+def set_forward_context(forward_batch: ForwardBatch, attention_layers: List[Any]):
+    global _forward_context
+    prev_forward_context = _forward_context
+    _forward_context = ForwardContext()
+    _forward_context.set_forward_batch(forward_batch)
+    _forward_context.set_attention_layers(attention_layers)
+    try:
+        yield
+    finally:
+        _forward_context = prev_forward_context