add cudagraph.

lmdeploy/pytorch/backends/dlinfer/maca/graph_runner.py

@@ -0,0 +1,300 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import Any, Dict, List, Tuple
+
+import torch
+from torch import Tensor
+
+from lmdeploy.pytorch.config import BackendConfig, CacheConfig, ModelConfig
+from lmdeploy.pytorch.model_inputs import StepContext
+from lmdeploy.pytorch.models.utils.cudagraph import CudaGraphMeta
+from lmdeploy.utils import get_logger
+
+from ...graph_runner import GraphRunner
+
+logger = get_logger('lmdeploy')
+
+BuffType = Dict[str, Tensor]
+
+
+def round_up_to_multiple_of_8(n: int):
+    return (n + 7) // 8 * 8
+
+
+def _false(*args, **kwargs):
+    """default value of not support cuda graph."""
+    return False
+
+
+class MACASingleGraphRunner:
+    """MACA single graph runner."""
+
+    def __init__(
+        self,
+        model: torch.nn.Module,
+        max_batches: int,
+        max_tokens: int,
+        num_blocks: int,
+        is_decoding: bool,
+        pool: Tuple[int, int],
+        device: torch.device,
+    ):
+        self.model = model
+        self.ctx_mgr = model.ctx_mgr
+        self.meta = CudaGraphMeta(
+            max_batchs=max_batches,
+            max_tokens=max_tokens,
+            num_blocks=num_blocks,
+            is_decoding=is_decoding,
+            device=device,
+            input_buffers=dict(),
+            output_buffers=dict(),
+        )
+        self.device = device
+        self.max_batches = max_batches
+        self.max_tokens = max_tokens
+        self.num_blocks = num_blocks
+        self.is_decoding = is_decoding
+        self.pool = pool
+        self._graph: torch.cuda.CUDAGraph = None
+
+    def capture(self, **kwargs):
+        """capture graph."""
+        self.meta.input_buffers = self.make_buffers_cudagraph(
+            self.meta, **kwargs)
+        padded_kwargs = self.fill_buffers_cudagraph(self.meta, **kwargs)
+        context = self.ctx_mgr.current_context()
+        self.update_context_cudagraph(self.meta, context)
+        current_stream = torch.cuda.current_stream()
+
+        output = self.model(**padded_kwargs)
+
+        # warmup
+        self._graph = torch.cuda.CUDAGraph()
+        # unsafe kernel call in other thread might invalid the capture
+        # so we set thread_safe capture mode here.
+        with torch.cuda.graph(self._graph,
+                              pool=self.pool,
+                              stream=current_stream,
+                              capture_error_mode='thread_local'):
+            output = self.model(**padded_kwargs)
+
+        output_buffers = dict(logits=output)
+        self.meta.output_buffers = output_buffers
+        return output
+
+    def forward(self, **kwargs):
+        """forward."""
+        num_tokens = kwargs['input_ids'].size(-1)
+        assert self._graph is not None
+        self.fill_buffers_cudagraph(self.meta, **kwargs)
+        context = self.ctx_mgr.current_context()
+        self.update_context_cudagraph(self.meta, context)
+
+        self._graph.replay()
+        output = self.meta.output_buffers['logits'][:, :num_tokens]
+        return output
+
+    def make_buffers_cudagraph(self, graph_meta: CudaGraphMeta, *args,
+                               **kwargs) -> BuffType:
+        """make cudagraph buffers from forward inputs."""
+        max_batches = graph_meta.max_batchs
+        max_tokens = graph_meta.max_tokens
+        num_blocks = graph_meta.num_blocks
+        device = graph_meta.device
+
+        input_buffers: BuffType = dict()
+        input_buffers['input_ids'] = torch.zeros(1,
+                                                 max_tokens,
+                                                 dtype=torch.int32,
+                                                 device=device)
+        input_buffers['position_ids'] = torch.ones((1, max_tokens),
+                                                   dtype=torch.int32,
+                                                   device=device)
+
+        input_buffers['block_offsets'] = torch.zeros((max_batches, num_blocks),
+                                                     dtype=torch.int32,
+                                                     device=device)
+
+        input_buffers['q_start_loc'] = torch.arange(max_batches + 1,
+                                                    dtype=torch.int32,
+                                                    device=device)
+
+        input_buffers['q_seqlens'] = torch.ones(max_batches,
+                                                dtype=torch.int32,
+                                                device=device)
+
+        input_buffers['kv_seqlens'] = torch.ones(max_batches,
+                                                 dtype=torch.int32,
+                                                 device=device)
+
+        input_buffers['kv_start_indices'] = torch.ones((max_batches, 1),
+                                                       dtype=torch.int64,
+                                                       device=device)
+
+        input_buffers['local_adapter_ids'] = torch.zeros(max_batches,
+                                                         dtype=torch.int32,
+                                                         device=device)
+        return input_buffers
+
+    def fill_buffers_cudagraph(self, graph_meta: CudaGraphMeta,
+                               input_ids: Tensor, position_ids: Tensor,
+                               past_key_values: List, attn_metadata: Any,
+                               inputs_embeds: Tensor,
+                               **kwargs) -> Dict[str, Tensor]:
+        """fill cudagraph buffers from forward inputs."""
+        is_decoding = graph_meta.is_decoding
+        block_offsets: Tensor = attn_metadata.block_offsets
+        q_start_loc: Tensor = attn_metadata.q_start_loc
+        q_seqlens: Tensor = attn_metadata.q_seqlens
+        kv_seqlens: Tensor = attn_metadata.kv_seqlens
+        kv_start_indices: Tensor = attn_metadata.kv_start_indices
+
+        input_buffers: BuffType = graph_meta.input_buffers
+
+        batch_size, num_blocks = block_offsets.size()
+        num_tokens = input_ids.size(-1)
+        q_start_loc_size = q_start_loc.size(0)
+
+        # fill buffer
+        input_buffers['input_ids'][:, :num_tokens] = input_ids
+        input_buffers['position_ids'][:, :num_tokens] = position_ids
+        input_buffers[
+            'block_offsets'][:batch_size, :num_blocks] = block_offsets
+        input_buffers['q_seqlens'][:batch_size] = q_seqlens
+        input_buffers['kv_seqlens'][:batch_size] = kv_seqlens
+
+        input_buffers['q_start_loc'][:q_start_loc_size] = q_start_loc
+        input_buffers['kv_start_indices'][:batch_size] = kv_start_indices
+
+        if inputs_embeds is not None:
+            emb_size = inputs_embeds.size(-1)
+            if 'inputs_embeds' not in input_buffers:
+                max_num_tokens = input_buffers['input_ids'].size(-1)
+                input_buffers['inputs_embeds'] = inputs_embeds.new_zeros(
+                    1, max_num_tokens, emb_size)
+            input_buffers['inputs_embeds'][:, :num_tokens] = inputs_embeds
+
+        # create inputs
+        new_batch_size = round_up_to_multiple_of_8(batch_size)
+        q_start_loc_size = round_up_to_multiple_of_8(q_start_loc_size)
+
+        attn_metadata.block_offsets = input_buffers[
+            'block_offsets'][:new_batch_size]
+        attn_metadata.q_start_loc = input_buffers[
+            'q_start_loc'][:q_start_loc_size]
+        attn_metadata.q_seqlens = input_buffers['q_seqlens'][:new_batch_size]
+        attn_metadata.kv_seqlens = input_buffers['kv_seqlens'][:new_batch_size]
+        attn_metadata.kv_start_indices = input_buffers[
+            'kv_start_indices'][:new_batch_size]
+
+        new_inputs = dict(
+            past_key_values=past_key_values,
+            attn_metadata=attn_metadata,
+        )
+
+        if is_decoding:
+            new_inputs['input_ids'] = input_buffers[
+                'input_ids'][:, :new_batch_size]
+            new_inputs['position_ids'] = input_buffers[
+                'position_ids'][:, :new_batch_size]
+        else:
+            new_inputs['input_ids'] = input_buffers['input_ids']
+            new_inputs['position_ids'] = input_buffers['position_ids']
+
+        if inputs_embeds is not None:
+            if is_decoding:
+                new_inputs['inputs_embeds'] = input_buffers[
+                    'inputs_embeds'][:, :new_batch_size]
+            else:
+                new_inputs['inputs_embeds'] = input_buffers['inputs_embeds']
+
+        new_inputs.update(kwargs)
+        return new_inputs
+
+    def update_context_cudagraph(self, graph_meta, context):
+        """update step context with input buffers."""
+        input_buffers = graph_meta.input_buffers
+        context.q_seqlens = input_buffers['q_seqlens']
+        context.kv_seqlens = input_buffers['kv_seqlens']
+        context.q_start_loc = input_buffers['q_start_loc']
+        context.kv_start_indices = input_buffers['kv_start_indices']
+
+    def __del__(self):
+        """del."""
+        del self._graph
+
+
+class MACAGraphRunner(GraphRunner):
+    """MACA graph runner."""
+
+    def __init__(self, model: torch.nn.Module, model_config: ModelConfig,
+                 cache_config: CacheConfig, backend_config: BackendConfig,
+                 device: torch.device):
+        super().__init__(model, model_config, cache_config, backend_config,
+                         device)
+        self.max_batches = cache_config.max_batches
+        self.max_tokens = cache_config.max_prefill_token_num
+        self.num_blocks = cache_config.num_gpu_blocks
+
+        self.enable_graph = self.check_enable_graph()
+
+        self.graph_pool_handle = torch.cuda.graph_pool_handle()
+        self._runner_map: Dict[Any, MACASingleGraphRunner] = dict()
+
+    def check_enable_graph(self):
+        """check enable graph."""
+        if self.backend_config.eager_mode:
+            return _false
+
+        return getattr(self.model, 'support_cuda_graph', _false)
+
+    def get_graph_key(self, input_ids: torch.Tensor,
+                      position_ids: torch.Tensor, past_key_values: List,
+                      attn_metadata: Any, inputs_embeds: torch.Tensor,
+                      **kwargs):
+        """get graph key."""
+        context = self.ctx_mgr.current_context()
+        is_decoding = context.is_decoding
+        num_tokens = input_ids.numel()
+        new_num_tokens = round_up_to_multiple_of_8(num_tokens)
+        return (new_num_tokens, is_decoding)
+
+    def __call__(self, **kwargs):
+        """call."""
+        enable_graph = self.enable_graph(**kwargs)
+        graph_key = self.get_graph_key(**kwargs)
+        max_tokens = graph_key[0]
+        is_decoding = graph_key[1]
+
+        if (not enable_graph) or (not is_decoding):
+            return self.model(**kwargs)
+
+        if graph_key not in self._runner_map:
+            max_batches = max_tokens if is_decoding else self.max_batches
+            runner = MACASingleGraphRunner(self.model,
+                                           max_batches=max_batches,
+                                           max_tokens=max_tokens,
+                                           num_blocks=self.num_blocks,
+                                           is_decoding=is_decoding,
+                                           pool=self.graph_pool_handle,
+                                           device=self.device)
+            runner.capture(**kwargs)
+            self._runner_map[graph_key] = runner
+        else:
+            runner = self._runner_map[graph_key]
+
+        output = runner.forward(**kwargs)
+        return output
+
+    def prepare_inputs_for_generation(
+        self,
+        past_key_values: List[List[torch.Tensor]],
+        inputs_embeds: torch.Tensor = None,
+        context: StepContext = None,
+    ):
+        """prepare inputs."""
+        return self.model.prepare_inputs_for_generation(
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            context=context,
+        )

lmdeploy/pytorch/backends/dlinfer/maca/op_backend.py

@@ -3,6 +3,7 @@
 
 import torch
 
+from lmdeploy.pytorch.config import BackendConfig, CacheConfig, ModelConfig
 from lmdeploy.utils import get_logger
 
 from ..op_backend import DlinferOpsBackend
@@ -105,3 +106,13 @@ def get_total_slots():
 
         step_context.attn_metadata = attn_metadata
         return step_context
+
+    @staticmethod
+    def build_graph_runner(model: torch.nn.Module, model_config: ModelConfig,
+                           cache_config: CacheConfig,
+                           backend_config: BackendConfig,
+                           device: torch.device):
+        """build graph runner."""
+        from .graph_runner import MACAGraphRunner
+        return MACAGraphRunner(model, model_config, cache_config,
+                               backend_config, device)