cache-dit 0.3.2__py3-none-any.whl → 1.0.14__py3-none-any.whl

cache_dit/caching/cache_blocks/pattern_3_4_5.py

@@ -0,0 +1,543 @@
+import torch
+
+from cache_dit.caching import ForwardPattern
+from cache_dit.caching.cache_contexts.cache_manager import (
+    ContextNotExistError,
+)
+from cache_dit.caching.cache_blocks.pattern_base import (
+    CachedBlocks_Pattern_Base,
+)
+from cache_dit.caching.cache_contexts.prune_context import PrunedContext
+from cache_dit.caching.cache_contexts.prune_manager import (
+    PrunedContextManager,
+)
+from cache_dit.caching.cache_types import CacheType
+
+from cache_dit.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class CachedBlocks_Pattern_3_4_5(CachedBlocks_Pattern_Base):
+    _supported_patterns = [
+        ForwardPattern.Pattern_3,
+        ForwardPattern.Pattern_4,
+        ForwardPattern.Pattern_5,
+    ]
+
+    def call_blocks(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        # Call all blocks to process the hidden states without cache.
+        new_encoder_hidden_states = None
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+            hidden_states, new_encoder_hidden_states = (
+                self._process_block_outputs(hidden_states)
+            )
+
+        return hidden_states, new_encoder_hidden_states
+
+    @torch.compiler.disable
+    def _process_block_outputs(
+        self, hidden_states: torch.Tensor | tuple
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        # Process the outputs for the block.
+        new_encoder_hidden_states = None
+        if not isinstance(hidden_states, torch.Tensor):  # Pattern 4, 5
+            if len(hidden_states) == 2:
+                if isinstance(hidden_states[1], torch.Tensor):
+                    hidden_states, new_encoder_hidden_states = hidden_states
+                    if not self.forward_pattern.Return_H_First:
+                        hidden_states, new_encoder_hidden_states = (
+                            new_encoder_hidden_states,
+                            hidden_states,
+                        )
+                elif isinstance(hidden_states[0], torch.Tensor):
+                    hidden_states = hidden_states[0]
+                else:
+                    raise ValueError("Unexpected hidden_states format.")
+            else:
+                assert (
+                    len(hidden_states) == 1
+                ), f"Unexpected output length: {len(hidden_states)}"
+                hidden_states = hidden_states[0]
+        return hidden_states, new_encoder_hidden_states
+
+    @torch.compiler.disable
+    def _process_forward_outputs(
+        self,
+        hidden_states: torch.Tensor,
+        new_encoder_hidden_states: torch.Tensor | None,
+    ) -> (
+        torch.Tensor
+        | tuple[torch.Tensor, torch.Tensor]
+        | tuple[torch.Tensor, None]
+    ):
+        if self.forward_pattern.Return_H_Only:
+            return hidden_states
+        else:
+            if self.forward_pattern.Return_H_First:
+                return (hidden_states, new_encoder_hidden_states)
+            else:
+                return (new_encoder_hidden_states, hidden_states)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        # Use it's own cache context.
+        try:
+            self.context_manager.set_context(self.cache_context)
+            self._check_cache_params()
+        except ContextNotExistError as e:
+            logger.warning(f"context not exist: {e}, skip cache.")
+            hidden_states, new_encoder_hidden_states = self.call_blocks(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+            return self._process_forward_outputs(
+                hidden_states, new_encoder_hidden_states
+            )
+
+        original_hidden_states = hidden_states
+        # Call first `n` blocks to process the hidden states for
+        # more stable diff calculation.
+        hidden_states, new_encoder_hidden_states = self.call_Fn_blocks(
+            hidden_states,
+            *args,
+            **kwargs,
+        )
+
+        Fn_hidden_states_residual = self._get_Fn_residual(
+            original_hidden_states, hidden_states
+        )
+        del original_hidden_states
+
+        self.context_manager.mark_step_begin()
+        # Residual L1 diff or Hidden States L1 diff
+        can_use_cache = self.context_manager.can_cache(
+            (
+                Fn_hidden_states_residual
+                if not self.context_manager.is_l1_diff_enabled()
+                else hidden_states
+            ),
+            parallelized=self._is_parallelized(),
+            prefix=(
+                f"{self.cache_prefix}_Fn_residual"
+                if not self.context_manager.is_l1_diff_enabled()
+                else f"{self.cache_prefix}_Fn_hidden_states"
+            ),
+        )
+
+        torch._dynamo.graph_break()
+        if can_use_cache:
+            self.context_manager.add_cached_step()
+            del Fn_hidden_states_residual
+            hidden_states, new_encoder_hidden_states = (
+                self.context_manager.apply_cache(
+                    hidden_states,
+                    new_encoder_hidden_states,  # encoder_hidden_states not use cache
+                    prefix=(
+                        f"{self.cache_prefix}_Bn_residual"
+                        if self.context_manager.is_cache_residual()
+                        else f"{self.cache_prefix}_Bn_hidden_states"
+                    ),
+                    encoder_prefix=(
+                        f"{self.cache_prefix}_Bn_residual"
+                        if self.context_manager.is_encoder_cache_residual()
+                        else f"{self.cache_prefix}_Bn_hidden_states"
+                    ),
+                )
+            )
+            torch._dynamo.graph_break()
+            # Call last `n` blocks to further process the hidden states
+            # for higher precision.
+            if self.context_manager.Bn_compute_blocks() > 0:
+                hidden_states, new_encoder_hidden_states = self.call_Bn_blocks(
+                    hidden_states,
+                    *args,
+                    **kwargs,
+                )
+        else:
+            self.context_manager.set_Fn_buffer(
+                Fn_hidden_states_residual,
+                prefix=f"{self.cache_prefix}_Fn_residual",
+            )
+            if self.context_manager.is_l1_diff_enabled():
+                # for hidden states L1 diff
+                self.context_manager.set_Fn_buffer(
+                    hidden_states,
+                    f"{self.cache_prefix}_Fn_hidden_states",
+                )
+            del Fn_hidden_states_residual
+            torch._dynamo.graph_break()
+            old_encoder_hidden_states = new_encoder_hidden_states
+            (
+                hidden_states,
+                new_encoder_hidden_states,
+                hidden_states_residual,
+            ) = self.call_Mn_blocks(  # middle
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+
+            torch._dynamo.graph_break()
+            if self.context_manager.is_cache_residual():
+                self.context_manager.set_Bn_buffer(
+                    hidden_states_residual,
+                    prefix=f"{self.cache_prefix}_Bn_residual",
+                )
+            else:
+                self.context_manager.set_Bn_buffer(
+                    hidden_states,
+                    prefix=f"{self.cache_prefix}_Bn_hidden_states",
+                )
+
+            if new_encoder_hidden_states is not None:
+                new_encoder_hidden_states_residual = (
+                    new_encoder_hidden_states - old_encoder_hidden_states
+                )
+            if self.context_manager.is_encoder_cache_residual():
+                if new_encoder_hidden_states is not None:
+                    self.context_manager.set_Bn_encoder_buffer(
+                        new_encoder_hidden_states_residual,
+                        prefix=f"{self.cache_prefix}_Bn_residual",
+                    )
+            else:
+                if new_encoder_hidden_states is not None:
+                    self.context_manager.set_Bn_encoder_buffer(
+                        new_encoder_hidden_states_residual,
+                        prefix=f"{self.cache_prefix}_Bn_hidden_states",
+                    )
+            torch._dynamo.graph_break()
+            # Call last `n` blocks to further process the hidden states
+            # for higher precision.
+            if self.context_manager.Bn_compute_blocks() > 0:
+                hidden_states, new_encoder_hidden_states = self.call_Bn_blocks(
+                    hidden_states,
+                    *args,
+                    **kwargs,
+                )
+
+        torch._dynamo.graph_break()
+
+        return self._process_forward_outputs(
+            hidden_states,
+            new_encoder_hidden_states,
+        )
+
+    def call_Fn_blocks(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        new_encoder_hidden_states = None
+        for block in self._Fn_blocks():
+            hidden_states = block(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+            hidden_states, new_encoder_hidden_states = (
+                self._process_block_outputs(hidden_states)
+            )
+
+        return hidden_states, new_encoder_hidden_states
+
+    def call_Mn_blocks(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        original_hidden_states = hidden_states
+        new_encoder_hidden_states = None
+        for block in self._Mn_blocks():
+            hidden_states = block(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+
+            hidden_states, new_encoder_hidden_states = (
+                self._process_block_outputs(hidden_states)
+            )
+
+        # compute hidden_states residual
+        hidden_states = hidden_states.contiguous()
+        hidden_states_residual = hidden_states - original_hidden_states.to(
+            hidden_states.device
+        )
+
+        return (
+            hidden_states,
+            new_encoder_hidden_states,
+            hidden_states_residual,
+        )
+
+    def call_Bn_blocks(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        new_encoder_hidden_states = None
+        if self.context_manager.Bn_compute_blocks() == 0:
+            return hidden_states, new_encoder_hidden_states
+
+        for block in self._Bn_blocks():
+            hidden_states = block(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+
+            hidden_states, new_encoder_hidden_states = (
+                self._process_block_outputs(hidden_states)
+            )
+
+        return hidden_states, new_encoder_hidden_states
+
+
+class PrunedBlocks_Pattern_3_4_5(CachedBlocks_Pattern_3_4_5):
+    _supported_patterns = [
+        ForwardPattern.Pattern_3,
+        ForwardPattern.Pattern_4,
+        ForwardPattern.Pattern_5,
+    ]
+    pruned_blocks_step: int = 0  # number of pruned blocks in current step
+
+    def __init__(
+        self,
+        # 0. Transformer blocks configuration
+        transformer_blocks: torch.nn.ModuleList,
+        transformer: torch.nn.Module = None,
+        forward_pattern: ForwardPattern = ForwardPattern.Pattern_0,
+        check_forward_pattern: bool = True,
+        check_num_outputs: bool = True,
+        # 1. Prune context configuration
+        cache_prefix: str = None,  # maybe un-need.
+        cache_context: PrunedContext | str = None,
+        context_manager: PrunedContextManager = None,
+        cache_type: CacheType = CacheType.DBPrune,
+        **kwargs,
+    ):
+        super().__init__(
+            # 0. Transformer blocks configuration
+            transformer_blocks,
+            transformer=transformer,
+            forward_pattern=forward_pattern,
+            check_forward_pattern=check_forward_pattern,
+            check_num_outputs=check_num_outputs,
+            # 1. Cache context configuration
+            cache_prefix=cache_prefix,
+            cache_context=cache_context,
+            context_manager=context_manager,
+            cache_type=cache_type,
+            **kwargs,
+        )
+        assert isinstance(
+            self.context_manager, PrunedContextManager
+        ), "context_manager must be PrunedContextManager for PrunedBlocks."
+        self.context_manager: PrunedContextManager = (
+            self.context_manager
+        )  # For type hint
+
+    @torch.compiler.disable
+    def _check_cache_type(self):
+        assert (
+            self.cache_type == CacheType.DBPrune
+        ), f"Cache type {self.cache_type} is not supported for PrunedBlocks."
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ):
+        self.pruned_blocks_step: int = 0  # reset for each step
+
+        # Use it's own cache context.
+        try:
+            self.context_manager.set_context(self.cache_context)
+            self._check_cache_params()
+        except ContextNotExistError as e:
+            logger.warning(f"context not exist: {e}, skip prune.")
+            hidden_states, new_encoder_hidden_states = self.call_blocks(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+            return self._process_forward_outputs(
+                hidden_states, new_encoder_hidden_states
+            )
+
+        self.context_manager.mark_step_begin()
+
+        if self._check_if_context_parallel_enabled(self.transformer_blocks[0]):
+            raise RuntimeError(
+                "Block level Context parallelism is not supported in PrunedBlocks."
+            )
+
+        # Call all blocks with prune strategy to process the hidden states.
+        new_encoder_hidden_states = None
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states, new_encoder_hidden_states = self.compute_or_prune(
+                i,
+                block,
+                hidden_states,
+                new_encoder_hidden_states,
+                *args,
+                **kwargs,
+            )
+
+        self.context_manager.add_pruned_block(self.pruned_blocks_step)
+        self.context_manager.add_actual_block(self.num_blocks)
+
+        return self._process_forward_outputs(
+            hidden_states,
+            new_encoder_hidden_states,
+        )
+
+    @property
+    @torch.compiler.disable
+    def num_blocks(self):
+        return len(self.transformer_blocks)
+
+    @torch.compiler.disable
+    def _skip_prune(self, block_id: int) -> bool:
+        # Wrap for non compiled mode.
+        return block_id in self.context_manager.get_non_prune_blocks_ids(
+            self.num_blocks
+        )
+
+    @torch.compiler.disable
+    def _maybe_prune(
+        self,
+        block_id: int,  # Block index in the transformer blocks
+        hidden_states: torch.Tensor,  # hidden_states or residual
+        prefix: str = "Bn_original",  # prev step name for single blocks
+    ):
+        # Wrap for non compiled mode.
+        can_use_prune = False
+        if not self._skip_prune(block_id):
+            can_use_prune = self.context_manager.can_prune(
+                hidden_states,  # curr step
+                parallelized=self._is_parallelized(),
+                prefix=prefix,  # prev step
+            )
+        self.pruned_blocks_step += int(can_use_prune)
+        return can_use_prune
+
+    def compute_or_prune(
+        self,
+        block_id: int,  # Block index in the transformer blocks
+        # Below are the inputs to the block
+        block,  # The transformer block to be executed
+        hidden_states: torch.Tensor,
+        new_encoder_hidden_states: torch.Tensor | None,
+        *args,
+        **kwargs,
+    ):
+        original_hidden_states = hidden_states
+        original_encoder_hidden_states = new_encoder_hidden_states
+
+        can_use_prune = self._maybe_prune(
+            block_id,
+            hidden_states,
+            prefix=f"{self.cache_prefix}_{block_id}_Fn_original",
+        )
+
+        # Prune steps: Prune current block and reuse the cached
+        # residuals for hidden states approximate.
+        torch._dynamo.graph_break()
+        if can_use_prune:
+            self.context_manager.add_pruned_step()
+            hidden_states, new_encoder_hidden_states = (
+                self.context_manager.apply_prune(
+                    hidden_states,
+                    new_encoder_hidden_states,
+                    prefix=(
+                        f"{self.cache_prefix}_{block_id}_Bn_residual"
+                        if self.context_manager.is_cache_residual()
+                        else f"{self.cache_prefix}_{block_id}_Bn_hidden_states"
+                    ),
+                    encoder_prefix=(
+                        f"{self.cache_prefix}_{block_id}_Bn_encoder_residual"
+                        if self.context_manager.is_encoder_cache_residual()
+                        else f"{self.cache_prefix}_{block_id}_Bn_encoder_hidden_states"
+                    ),
+                )
+            )
+            torch._dynamo.graph_break()
+        else:
+            # Normal steps: Compute the block and cache the residuals.
+            hidden_states = block(
+                hidden_states,
+                *args,
+                **kwargs,
+            )
+            hidden_states, new_encoder_hidden_states = (
+                self._process_block_outputs(
+                    hidden_states, new_encoder_hidden_states
+                )
+            )
+            if not self._skip_prune(block_id):
+                hidden_states = hidden_states.contiguous()
+                hidden_states_residual = hidden_states - original_hidden_states
+
+                if (
+                    new_encoder_hidden_states is not None
+                    and original_encoder_hidden_states is not None
+                ):
+                    new_encoder_hidden_states = (
+                        new_encoder_hidden_states.contiguous()
+                    )
+                    new_encoder_hidden_states_residual = (
+                        new_encoder_hidden_states
+                        - original_encoder_hidden_states
+                    )
+                else:
+                    new_encoder_hidden_states_residual = None
+
+                self.context_manager.set_Fn_buffer(
+                    original_hidden_states,
+                    prefix=f"{self.cache_prefix}_{block_id}_Fn_original",
+                )
+                if self.context_manager.is_cache_residual():
+                    self.context_manager.set_Bn_buffer(
+                        hidden_states_residual,
+                        prefix=f"{self.cache_prefix}_{block_id}_Bn_residual",
+                    )
+                else:
+                    self.context_manager.set_Bn_buffer(
+                        hidden_states,
+                        prefix=f"{self.cache_prefix}_{block_id}_Bn_hidden_states",
+                    )
+                if new_encoder_hidden_states_residual is not None:
+                    if self.context_manager.is_encoder_cache_residual():
+                        self.context_manager.set_Bn_encoder_buffer(
+                            new_encoder_hidden_states_residual,
+                            prefix=f"{self.cache_prefix}_{block_id}_Bn_encoder_residual",
+                        )
+                    else:
+                        self.context_manager.set_Bn_encoder_buffer(
+                            new_encoder_hidden_states_residual,
+                            prefix=f"{self.cache_prefix}_{block_id}_Bn_encoder_hidden_states",
+                        )
+            torch._dynamo.graph_break()
+
+        return hidden_states, new_encoder_hidden_states