Fix _speculative_sampling implementation (#28508)

src/transformers/generation/candidate_generator.py

@@ -171,12 +171,16 @@ def get_candidates(self, input_ids: torch.LongTensor) -> Tuple[torch.LongTensor,
         """
         input_ids = input_ids.to(self.assistant_model.device)
 
+        # Don't generate more than `max_length - 1` candidates since the target model generates one extra token.
+        new_cur_len = input_ids.shape[-1]
+        max_new_tokens = min(int(self.num_assistant_tokens), self.generation_config.max_length - new_cur_len - 1)
+        if max_new_tokens == 0:
+            return input_ids, None
+
         # 1. If it is not the first round of candidate generation, prepare the inputs based on the input_ids length
         # (which implicitly contains the number of accepted candidates from the previous round)
         has_past_key_values = self.assistant_kwargs.get("past_key_values", None) is not None
         if has_past_key_values:
-            new_cur_len = input_ids.shape[-1]
-
             new_cache_size = new_cur_len - 1
             self.assistant_kwargs["past_key_values"] = _crop_past_key_values(
                 self.assistant_model, self.assistant_kwargs["past_key_values"], new_cache_size - 1
@@ -190,7 +194,7 @@ def get_candidates(self, input_ids: torch.LongTensor) -> Tuple[torch.LongTensor,
         # 2. Forecast next N tokens using the assistant model.
         assistant_generation_kwargs = {
             self.input_ids_key: input_ids,
-            "max_new_tokens": int(self.num_assistant_tokens),
+            "max_new_tokens": max_new_tokens,
             "generation_config": self.generation_config,
             "logits_processor": self.logits_processor,
         }

src/transformers/generation/utils.py

@@ -4395,7 +4395,7 @@ def assisted_decoding(
                 else:
                     selected_tokens = new_logits.argmax(dim=-1)
 
-                candidate_new_tokens = candidate_input_ids[:, -candidate_length:]
+                candidate_new_tokens = candidate_input_ids[:, cur_len:]
                 n_matches = ((~(candidate_new_tokens == selected_tokens[:, :-1])).cumsum(dim=-1) < 1).sum()
 
                 # Ensure we don't generate beyond max_len or an EOS token
@@ -4531,41 +4531,47 @@ def _speculative_sampling(
 
     NOTE: Unless otherwise stated, the variable names match those in the paper.
     """
+    new_candidate_input_ids = candidate_input_ids[:, -candidate_length:]
     # Gets the probabilities from the logits. q_i and p_i denote the assistant and model probabilities of the tokens
     # selected by the assistant, respectively.
     q = candidate_logits.softmax(dim=-1)
-    q_i = q[:, torch.arange(candidate_length), candidate_input_ids[:, -candidate_length:]].squeeze(0, 1)
+    q_i = q[:, torch.arange(candidate_length), new_candidate_input_ids].squeeze(0, 1)
     p = new_logits.softmax(dim=-1)
-    p_i = p[:, torch.arange(candidate_length), candidate_input_ids[:, -candidate_length:]].squeeze(0, 1)
+    p_i = p[:, torch.arange(candidate_length), new_candidate_input_ids].squeeze(0, 1)
     probability_ratio = p_i / q_i
 
     # When probability_ratio > 1 (i.e. q_i(x) < p_i(x), or "assistant probability of the candidate token is smaller
     # than the model probability for the same token"), keep the token. Otherwise reject with p = 1 - probability_ratio
     # (= keep with p = probability_ratio). Keep all the tokens until the first rejection
     r_i = torch.rand_like(probability_ratio)
     is_accepted = r_i <= probability_ratio
-    n_matches = (~is_accepted.cumsum(dim=-1) < 1).sum()  # this is `n` in algorithm 1
+    n_matches = ((~is_accepted).cumsum(dim=-1) < 1).sum()  # this is `n` in algorithm 1
 
     # Ensure we don't generate beyond max_len or an EOS token (not in algorithm 1, but needed for correct behavior)
     if last_assistant_token_is_eos and n_matches == candidate_length:
+        # Output length is assumed to be `n_matches + 1`. Since we won't generate another token with the target model
+        # due to acceptance on EOS we fix `n_matches`
         n_matches -= 1
-    n_matches = min(n_matches, max_matches)
-
-    # Next token selection: if there is a rejection, adjust the distribution from the main model before sampling.
-    gamma = candidate_logits.shape[1]
-    p_n_plus_1 = p[:, n_matches, :]
-    if n_matches < gamma:
-        q_n_plus_1 = q[:, n_matches, :]
-        p_prime = torch.clamp((p_n_plus_1 - q_n_plus_1), min=0).softmax(dim=-1)
+        valid_tokens = new_candidate_input_ids[:, : n_matches + 1]
     else:
-        p_prime = p_n_plus_1
-    t = torch.multinomial(p_prime, num_samples=1).squeeze(1)[None, :]
+        n_matches = min(n_matches, max_matches)
+
+        # Next token selection: if there is a rejection, adjust the distribution from the main model before sampling.
+        gamma = min(candidate_logits.shape[1], max_matches)
+        p_n_plus_1 = p[:, n_matches, :]
+        if n_matches < gamma:
+            q_n_plus_1 = q[:, n_matches, :]
+            p_prime = torch.clamp((p_n_plus_1 - q_n_plus_1), min=0)
+            p_prime.div_(p_prime.sum())
+        else:
+            p_prime = p_n_plus_1
+        t = torch.multinomial(p_prime, num_samples=1).squeeze(1)[None, :]
 
-    # The selected tokens include the matches (if any) plus the next sampled tokens
-    if n_matches > 0:
-        valid_tokens = torch.cat((candidate_input_ids[:, -n_matches:], t), dim=-1)
-    else:
-        valid_tokens = t
+        # The selected tokens include the matches (if any) plus the next sampled tokens
+        if n_matches > 0:
+            valid_tokens = torch.cat((new_candidate_input_ids[:, :n_matches], t), dim=-1)
+        else:
+            valid_tokens = t
 
     return valid_tokens, n_matches
 

tests/generation/test_utils.py

@@ -88,6 +88,7 @@
         TopKLogitsWarper,
         TopPLogitsWarper,
     )
+    from transformers.generation.utils import _speculative_sampling
 
 
 class GenerationTesterMixin:
@@ -2391,6 +2392,43 @@ def test_top_k_top_p_filtering_with_filter_value(self):
 
         self.assertTrue(torch.allclose(expected_output, output, atol=1e-12))
 
+    def test_speculative_sampling(self):
+        # assume vocab size 10, input length 5 + 3 generated candidates
+        candidate_input_ids = torch.tensor([[8, 0, 3, 9, 8, 1, 4, 5]])  # input tokens
+        candidate_logits = torch.tensor(
+            [
+                [
+                    [-10.0, 10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0],  # generated 1
+                    [-10.0, -10.0, -10.0, -10.0, 10.0, -10.0, -10.0, -10.0, -10.0, -10.0],  # generated 4
+                    [-10.0, -10.0, -10.0, -10.0, -10.0, 10.0, -10.0, -10.0, -10.0, -10.0],  # generated 5
+                ]
+            ]
+        )
+        candidate_length = 3
+        inf = float("inf")
+        new_logits = torch.tensor(
+            [
+                [
+                    [-10.0, 10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0],  # accepts 1
+                    [-10.0, -10.0, -10.0, -10.0, 10.0, -10.0, -10.0, -10.0, -10.0, -10.0],  # accepts 4
+                    [-inf, -inf, -inf, -inf, -inf, -inf, -inf, -inf, 10.0, -inf],  # rejects 5, accepts 8
+                    [-10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0, -10.0],  # N/A
+                ]
+            ]
+        )
+        last_assistant_token_is_eos = False
+        max_matches = 5
+        validated_tokens, n_matches = _speculative_sampling(
+            candidate_input_ids,
+            candidate_logits,
+            candidate_length,
+            new_logits,
+            last_assistant_token_is_eos,
+            max_matches,
+        )
+        self.assertTrue(n_matches.item() == 2)
+        self.assertTrue(validated_tokens.tolist()[0] == [1, 4, 8])
+
 
 @require_torch
 class GenerationIntegrationTests(unittest.TestCase, GenerationIntegrationTestsMixin):