feat: apply dry logic

0xymoro · 0xymoro · commit 03ee5df60239 · 2025-04-15T16:34:01.000-07:00
diff --git a/vllm/v1/sample/ops/penalties.py b/vllm/v1/sample/ops/penalties.py
@@ -1,6 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import torch
+from typing import List, Set, Optional, Dict, Any
+from vllm.v1.sample.metadata import SamplingMetadata 
 
 from vllm.model_executor.layers.utils import apply_penalties
 from vllm.utils import is_pin_memory_available, make_tensor_with_pad
@@ -56,3 +58,188 @@ def _convert_to_tensors(output_token_ids: list[list[int]], vocab_size: int,
         pin_memory=is_pin_memory_available(),
     )
     return output_tokens_tensor.to(device, non_blocking=True)
+
+
+# Constants for DRY
+_DRY_MAX_NGRAM = 12
+_DRY_MAX_OCCURRENCES = 8
+_DRY_EARLY_EXIT_MATCH_LEN = 8
+
+# Default DRY parameter values
+_DRY_DEFAULT_MULTIPLIER = 0.0
+_DRY_DEFAULT_BASE = 1.0
+_DRY_DEFAULT_ALLOWED_LEN = 3
+_DRY_DEFAULT_RANGE = 1500 
+_DRY_DEFAULT_BREAKERS: Set[int] = set()
+
+
+def apply_dry(
+    logits: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+) -> torch.Tensor:
+    """
+    Applies DRY (Don't Repeat Yourself) penalty to logits based on parameters
+    found in sampling_metadata.extra_data for each request.
+
+    Modifies logits in-place or returns a modified tensor.
+
+    Expected keys in extra_data[irow] (if DRY is active):
+        - 'dry_multiplier' (float)
+        - 'dry_base' (float)
+        - 'dry_allowed_len' (int)
+        - 'dry_range' (int)
+        - 'dry_breakers' (List[int] or Set[int])
+    """
+    batch_size, vocab_size = logits.shape
+    device = logits.device
+
+    # Assume extra_data is a list of dicts, one per request
+    # Or potentially accessed via methods on sampling_metadata
+    # Adjust access pattern if sampling_metadata structure differs
+    if not hasattr(sampling_metadata, 'extra_data') or sampling_metadata.extra_data is None:
+         # If no extra_data field exists or is None, cannot apply DRY
+         return logits
+
+    # Check if any request might have DRY enabled (basic check)
+    # More robust check would involve iterating through extra_data first
+    has_potential_dry = any(
+        data and data.get('dry_multiplier', _DRY_DEFAULT_MULTIPLIER) > 0
+        for data in sampling_metadata.extra_data
+    )
+    if not has_potential_dry:
+         return logits
+
+    # --- Iterate through each request in the batch ---
+    for irow in range(batch_size):
+        # Ensure sampling_metadata has data for this row index
+        if irow >= len(sampling_metadata.extra_data):
+             # If metadata doesn't cover this row (shouldn't happen in normal flow), skip
+             continue
+        extra_data = sampling_metadata.extra_data[irow]
+        if not extra_data:
+            continue
+
+        # Get DRY parameters for this row, using defaults if missing
+        multiplier = float(extra_data.get('dry_multiplier', _DRY_DEFAULT_MULTIPLIER))
+        if multiplier <= 0.0:
+            continue # DRY not active for this row
+
+        base = float(extra_data.get('dry_base', _DRY_DEFAULT_BASE))
+        allowed_length = int(extra_data.get('dry_allowed_len', _DRY_DEFAULT_ALLOWED_LEN))
+        dry_range = int(extra_data.get('dry_range', _DRY_DEFAULT_RANGE))
+        breakers_input = extra_data.get('dry_breakers', _DRY_DEFAULT_BREAKERS)
+        breakers = set(breakers_input) if breakers_input else _DRY_DEFAULT_BREAKERS
+
+        # 1. Construct the token sequence for this row
+        # Get prompt tokens (handle potential padding if needed)
+        # Assuming prompt_token_ids is available and correctly indexed
+        # Need prompt_lens if prompt_token_ids is padded
+        prompt_len_attr = getattr(sampling_metadata, 'prompt_lens', None)
+        current_prompt_len = prompt_len_attr[irow] if prompt_len_attr and irow < len(prompt_len_attr) else None
+
+        # Ensure prompt_token_ids covers this row
+        if irow >= sampling_metadata.prompt_token_ids.shape[0]:
+             continue # Skip if prompt data isn't available for this row
+        prompt_tensor_row = sampling_metadata.prompt_token_ids[irow]
+
+        if current_prompt_len is not None:
+             current_prompt_tokens = prompt_tensor_row[:current_prompt_len].tolist()
+        else:
+             # If prompt_lens is not available, we cannot reliably determine
+             # the prompt sequence length. Log a warning or raise an error,
+             # or potentially skip DRY for this row if appropriate.
+             # For now, let's skip if length is unknown.
+             # Consider adding logging: logger.warning("prompt_lens not available...")
+             continue # Skip DRY for this row if prompt length is unknown
+
+        # Get output tokens for this row
+        # Ensure output_token_ids covers this row
+        if irow >= len(sampling_metadata.output_token_ids):
+             continue # Skip if output data isn't available
+        current_output_tokens = sampling_metadata.output_token_ids[irow]
+        token_seq_list = current_prompt_tokens + current_output_tokens
+
+        # 2. Apply range limit
+        if dry_range > 0 and len(token_seq_list) > dry_range:
+            token_seq_list = token_seq_list[-dry_range:]
+
+        seq_len = len(token_seq_list)
+        if seq_len < 2:
+            continue # Need at least 2 tokens
+
+        last_token = token_seq_list[-1]
+        if last_token in breakers:
+            continue
+
+        # Convert to tensor for efficient processing
+        token_seq_tensor = torch.tensor(token_seq_list, dtype=torch.long, device=device)
+
+        # 3. Build break mask on device
+        break_mask = torch.zeros(seq_len, dtype=torch.bool, device=device)
+        if breakers:
+            breaker_tensor = torch.tensor(list(breakers), dtype=torch.long, device=device)
+            # Use broadcasting for efficiency
+            break_mask = torch.any(token_seq_tensor.unsqueeze(1) == breaker_tensor, dim=1)
+
+
+        # 4. Determine max n-gram length possible from the end
+        max_ngram = 0
+        for offset in range(1, min(seq_len, _DRY_MAX_NGRAM + 1)):
+            check_idx = -offset - 1
+            if check_idx < -seq_len: break
+            if break_mask[check_idx]: break
+            max_ngram = offset
+
+        if max_ngram < allowed_length:
+            continue
+
+        # 5. Find previous occurrences of last_token
+        endpoint_indices = (token_seq_tensor == last_token).nonzero(as_tuple=True)[0]
+        if len(endpoint_indices) < 2: continue
+
+        endpoint_indices = endpoint_indices[endpoint_indices != seq_len - 1]
+        if len(endpoint_indices) == 0: continue
+
+        if len(endpoint_indices) > _DRY_MAX_OCCURRENCES:
+            endpoint_indices = endpoint_indices[-_DRY_MAX_OCCURRENCES:]
+
+        # 6. Calculate match lengths for potential next tokens
+        ngram_lens = torch.zeros(vocab_size, dtype=torch.int32, device=device)
+        found_early_exit_match = False
+
+        for idx_tensor in reversed(endpoint_indices):
+            idx = idx_tensor.item()
+            match_len = 0
+            for unwind in range(1, max_ngram + 1):
+                current_idx = idx - unwind
+                history_idx = seq_len - 1 - unwind
+                if current_idx < 0: break
+                # Check breaks using the precomputed mask
+                if break_mask[current_idx] or break_mask[history_idx]: break
+                if token_seq_tensor[current_idx] != token_seq_tensor[history_idx]: break
+                match_len = unwind
+
+            if match_len >= allowed_length: # Match length must meet minimum
+                next_tok_idx = idx + 1
+                if next_tok_idx < seq_len:
+                    next_tok = token_seq_tensor[next_tok_idx].item()
+                    # Use match_len as the length of the *matched* sequence
+                    new_len = match_len
+                    current_max = ngram_lens[next_tok].item()
+                    ngram_lens[next_tok] = max(current_max, new_len)
+                    if new_len >= _DRY_EARLY_EXIT_MATCH_LEN:
+                        found_early_exit_match = True
+
+            if found_early_exit_match: break
+
+        # 7. Apply penalty to logits for this row
+        penalty_mask = ngram_lens > 0
+        if penalty_mask.any():
+            match_lengths_for_penalty = ngram_lens[penalty_mask]
+            # Clamp exponent >= 0
+            exponents = (match_lengths_for_penalty.float() - allowed_length).clamp_(min=0.0)
+            scales = base ** exponents
+            logits[irow, penalty_mask] -= multiplier * scales
+        # --- End of DRY logic for row ---
+
+    return logits
diff --git a/vllm/v1/sample/sampler.py b/vllm/v1/sample/sampler.py
@@ -8,7 +8,8 @@
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.ops.bad_words import apply_bad_words
 from vllm.v1.sample.ops.penalties import (apply_all_penalties,
-                                          apply_min_token_penalties)
+                                          apply_min_token_penalties,
+                                          apply_dry)
 from vllm.v1.sample.ops.topk_topp_sampler import TopKTopPSampler
 
 _SAMPLING_EPS = 1e-5
@@ -184,10 +185,13 @@ def apply_penalties(
         logits: torch.Tensor,
         sampling_metadata: SamplingMetadata,
     ) -> torch.Tensor:
+        # Apply min tokens penalty first (modifies logits in-place)
         if sampling_metadata.min_tokens:
             apply_min_token_penalties(logits,
                                       sampling_metadata.output_token_ids,
                                       sampling_metadata.min_tokens)
+
+        # Apply standard penalties (Freq, Presence, Repetition)
         if not sampling_metadata.no_penalties:
             assert sampling_metadata.prompt_token_ids is not None
             logits = apply_all_penalties(
@@ -196,8 +200,13 @@ def apply_penalties(
                 sampling_metadata.presence_penalties,
                 sampling_metadata.frequency_penalties,
                 sampling_metadata.repetition_penalties,
-                sampling_metadata.output_token_ids,
+                sampling_metadata.output_token_ids, # Passed as list of lists
             )
+
+        # Apply DRY penalty using the metadata object
+        # The check for whether DRY is active is now inside apply_dry
+        logits = apply_dry(logits, sampling_metadata)
+
         return logits
 
     def apply_min_p(
