Lookahead decoding eager implementation (#12491)

Summary:

Implement reference lookahead decoding for CoreML implementation.

Reviewed By: sxu

Differential Revision: D78323399

Author: viveknayakatmeta

examples/apple/coreml/llama/llama_transformer.py

@@ -6,6 +6,8 @@
 
 # Please refer to README.md in the same folder for more information.
 
+import logging
+from collections import defaultdict, deque
 from dataclasses import dataclass
 from functools import partial
 from typing import Dict, List, Optional, Tuple
@@ -23,6 +25,8 @@
 
 from torch import nn
 
+logger = logging.getLogger(__name__)
+
 
 def find_multiple(n: int, k: int) -> int:
     if n % k == 0:
@@ -507,6 +511,24 @@ def load_model(checkpoint_path, params_path, max_seq_length, use_cache_list):
 
 
 class InputManager:
+    class NGramCache:
+        def __init__(self, max_size: int):
+            self.cache = deque()
+            self.max_size = max_size
+
+        def add(self, ngram: List[int]):
+            if ngram in self.cache:
+                return
+            if len(self.cache) == self.max_size:
+                self.cache.popleft()
+            self.cache.append(ngram)
+
+        def __iter__(self):
+            return iter(self.cache)
+
+        def __str__(self):
+            return str(self.cache)
+
     def __init__(
         self,
         n_layers: int,
@@ -519,6 +541,7 @@ def __init__(
         dtype=torch.float16,
         minus_infinity=-torch.inf,
         cache_size=None,
+        lookahead_enabled: bool = False,
     ):
         if cache_size is None:
             cache_size = max_seq_length - seq_length
@@ -532,6 +555,8 @@ def __init__(
 
         self.seq_length = seq_length
         self.use_cache_list = use_cache_list
+        self.lookahead_enabled = lookahead_enabled
+        self.minus_infinity = minus_infinity
 
         if self.use_cache_list:
             self.k_caches = [
@@ -609,10 +634,10 @@ def _update_cache(self, start, length, new_k_caches, new_v_caches):
         if self.cache_pos == self.cache_size:
             self.cache_pos = 0
 
-    def update(self, input_length, new_k_caches, new_v_caches):
+    def update(self, input_length, new_k_caches, new_v_caches, update_pos=0):
         # Copy as much new cache data into cache as possible without wrapping
         amount_to_copy = min(input_length, self.cache_size - self.cache_pos)
-        self._update_cache(0, amount_to_copy, new_k_caches, new_v_caches)
+        self._update_cache(update_pos, amount_to_copy, new_k_caches, new_v_caches)
         if self.input_pos <= self.cache_size:
             self.attn_mask[:, (self.input_pos) : (self.input_pos + amount_to_copy)] = (
                 0.0
@@ -625,7 +650,7 @@ def update(self, input_length, new_k_caches, new_v_caches):
         )
         if remaining_to_copy > 0:
             self._update_cache(
-                amount_to_copy, remaining_to_copy, new_k_caches, new_v_caches
+                update_pos + amount_to_copy, remaining_to_copy, new_k_caches, new_v_caches
             )
 
         self.input_pos += input_length
@@ -661,3 +686,192 @@ def get_inputs_and_remaining_tokens(self, tokens: List[int]):
             self.get_inputs(tokens[0:processed_tokens]),
             tokens[processed_tokens:],
         )
+
+    def _get_lookahead_decoding_mask(
+        self, ngram_size: int, window_size: int, n_verifications: int
+    ) -> torch.Tensor:
+        mask = torch.full((self.seq_length, self.seq_length), self.minus_infinity)
+        mask[0][0] = 0.0
+
+        lookahead_submask = torch.triu(
+            torch.full((window_size, window_size), self.minus_infinity),
+            diagonal=1,
+        )
+        for i in range(ngram_size - 1):
+            offset = window_size * i
+            mask[offset : offset + window_size, :window_size] = lookahead_submask
+            for j in range(1, i + 1):
+                mask[
+                    offset : offset + window_size,
+                    window_size * j : window_size * (j + 1),
+                ].fill_diagonal_(0.0)
+
+        verification_offset = max(window_size * (ngram_size - 1), 1)
+        verification_submask = torch.triu(
+            torch.full((ngram_size - 1, ngram_size - 1), self.minus_infinity),
+            diagonal=1,
+        )
+        for i in range(n_verifications):
+            mask[
+                verification_offset + i * (ngram_size - 1) : verification_offset
+                + (i + 1) * (ngram_size - 1),
+                verification_offset + i * (ngram_size - 1) : verification_offset
+                + (i + 1) * (ngram_size - 1),
+            ] = verification_submask
+        mask[verification_offset:, :1] = 0.0
+
+        return mask
+
+    def _get_lookahead_position_offsets(
+        self, ngram_size: int, window_size: int, n_verifications: int
+    ) -> torch.Tensor:
+        pos_offsets = torch.zeros(self.seq_length, dtype=torch.int32)
+        idx = 0
+        if window_size > 0:
+            for i in range(ngram_size - 1):
+                for j in range(window_size):
+                    pos_offsets[idx] = i + j
+                    idx += 1
+        else:
+            pos_offsets[0] = 0
+            idx += 1
+
+        # Verification branches: [1, 2, ..., ngram_size - 1].
+        for _ in range(n_verifications):
+            for j in range(1, ngram_size):
+                pos_offsets[idx] = j
+                idx += 1
+
+        return pos_offsets
+
+    def lookahead_decode(
+        self,
+        model,
+        init_token: int,
+        n: int,
+        ngram_size: int,
+        window_size: int,
+        n_verifications: int,
+        stop_tokens: Optional[List[int]] = None,
+        ngram_caches: Optional[Dict[int, "InputManager.NGramCache"]] = None,
+    ) -> List[int]:
+        if not self.lookahead_enabled:
+            raise RuntimeError("Lookahead decoding is not enabled")
+
+        if (ngram_size - 1) * (window_size + n_verifications) > self.seq_length:
+            raise RuntimeError(
+                f"Lookahead decoding configuration not compatible with seq_length {self.seq_length}. "
+                f"Required: {(ngram_size - 1) * (window_size + n_verifications)}"
+            )
+
+        self.attn_mask[:, self.cache_size :] = self._get_lookahead_decoding_mask(
+            ngram_size, window_size, n_verifications
+        )
+        logger.debug("Lookahead decoding mask: ")
+        for i in range(self.seq_length):
+            logger.debug(
+                " ".join(
+                    ("X" if x == 0.0 else " ")
+                    for x in self.attn_mask[i][self.cache_size :]
+                )
+            )
+
+        offsets = self._get_lookahead_position_offsets(
+            ngram_size, window_size, n_verifications
+        )
+
+        stop_tokens = stop_tokens or []
+        verification_offset = window_size * (ngram_size - 1)
+
+        if ngram_caches is None:
+            ngram_caches = defaultdict(lambda: InputManager.NGramCache(n_verifications))
+        new_tokens = [init_token]
+        x = [init_token] * self.seq_length
+        inference_count = 0
+
+        while len(new_tokens) < n + 1:
+            cache = ngram_caches[x[0]]
+            for i, ngram in enumerate(cache):
+                for j, token in enumerate(ngram):
+                    x[verification_offset + i * (ngram_size - 1) + j] = token
+
+            logits, new_k, new_v = model(
+                tokens=torch.tensor([x], dtype=torch.int64),
+                input_pos=torch.tensor([self.input_pos], dtype=torch.long),
+                k_caches=self.k_caches,
+                v_caches=self.v_caches,
+                attn_mask=self.attn_mask,
+                input_len=torch.tensor([len(x)], dtype=torch.long),
+                rope_indices=self.input_pos + offsets,
+            )
+            inference_count += 1
+
+            # Greedy only
+            y = logits[0].argmax(dim=-1).tolist()
+            new_tokens.append(y[0])
+            logger.debug(f"{self.input_pos}: x = {x[0]}, y = {y[0]}")
+            if new_tokens[-1] in stop_tokens:
+                break
+
+            # Collect new n-grams.
+            for i in range(window_size):
+                key = x[i]
+                suffix = []
+                for j in range(1, ngram_size - 1):
+                    suffix.append(x[i + j * window_size])
+                suffix.append(y[i + window_size * (ngram_size - 2)])
+                ngram_caches[key].add(suffix)
+
+            # Verification.
+            longest_match = []
+            matched_branch = None
+            for i in range(n_verifications):
+                match = [y[0]]
+                j = 0
+                # for j in range(ngram_size - 1):
+                while (
+                    j < ngram_size - 1
+                    and x[verification_offset + (ngram_size - 1) * i + j] == match[-1]
+                ):
+                    match.append(y[verification_offset + (ngram_size - 1) * i + j])
+                    j += 1
+                if len(match) - 1 > len(longest_match):
+                    longest_match = match[1:]
+                    matched_branch = i
+
+            if matched_branch is not None:
+                logger.debug(
+                    f"Matched {len(longest_match)} additional tokens from n-grams: {longest_match}"
+                )
+                for stop in stop_tokens:
+                    if stop in longest_match:
+                        longest_match = longest_match[: longest_match.index(stop) + 1]
+
+                new_tokens.extend(longest_match)
+                branch_offset = verification_offset + (ngram_size - 1) * matched_branch
+                self.update(
+                    input_length=len(longest_match),
+                    new_k_caches=new_k,
+                    new_v_caches=new_v,
+                    update_pos=branch_offset,
+                )
+            else:
+                self.update(input_length=1, new_k_caches=new_k, new_v_caches=new_v)
+
+            # Update lookahead branch.
+            for i in range(ngram_size - 2):
+                for j in range(window_size):
+                    x[window_size * i + j] = x[window_size * (i + 1) + j]
+            for j in range(window_size):
+                x[window_size * (ngram_size - 2) + j] = y[
+                    window_size * (ngram_size - 2) + j
+                ]
+
+            x[0] = new_tokens[-1]
+            if new_tokens[-1] in stop_tokens:
+                break
+
+        logger.info(
+            f"Generated {len(new_tokens) - 1} tokens with {inference_count} inference(s)."
+        )
+        return new_tokens