import torch from collections import deque import torch.nn.functional as F from common import COMPUTE_DTYPE class KVCache: """ KV Cache designed for Flash Attention 3's flash_attn_with_kvcache API. Key differences from FA2-style cache: - Tensors are (B, T, H, D) not (B, H, T, D) - FA3 updates the cache in-place during flash_attn_with_kvcache - Position tracked per batch element via cache_seqlens tensor """ def __init__(self, batch_size, num_heads, seq_len, head_dim, num_layers, device, dtype): self.batch_size = batch_size self.max_seq_len = seq_len self.n_layers = num_layers self.n_heads = num_heads self.head_dim = head_dim # Pre-allocate cache tensors: (n_layers, B, T, H, D) self.k_cache = torch.zeros(num_layers, batch_size, seq_len, num_heads, head_dim, device=device, dtype=dtype) self.v_cache = torch.zeros(num_layers, batch_size, seq_len, num_heads, head_dim, device=device, dtype=dtype) # Current sequence length per batch element (FA3 needs int32) self.cache_seqlens = torch.zeros(batch_size, dtype=torch.int32, device=device) # Previous token's normalized embedding for smear (set by model forward pass) self.prev_embedding = None def reset(self): """Reset cache to empty state.""" self.cache_seqlens.zero_() self.prev_embedding = None def get_pos(self): """Get current position (assumes all batch elements at same position).""" return self.cache_seqlens[0].item() def get_layer_cache(self, layer_idx): """Return (k_cache, v_cache) views for a specific layer.""" return self.k_cache[layer_idx], self.v_cache[layer_idx] def advance(self, num_tokens): """Advance the cache position by num_tokens.""" self.cache_seqlens += num_tokens def prefill(self, other): """ Copy cached KV from another cache into this one. Used when we do batch=1 prefill and then want to generate multiple samples in parallel. """ assert self.get_pos() == 0, "Cannot prefill a non-empty KV cache" assert self.n_layers == other.n_layers and self.n_heads == other.n_heads and self.head_dim == other.head_dim assert self.max_seq_len >= other.max_seq_len other_pos = other.get_pos() self.k_cache[:, :, :other_pos, :, :] = other.k_cache[:, :, :other_pos, :, :] self.v_cache[:, :, :other_pos, :, :] = other.v_cache[:, :, :other_pos, :, :] self.cache_seqlens.fill_(other_pos) # Copy smear state: expand batch=1 prev_embedding to num_samples if other.prev_embedding is not None: self.prev_embedding = other.prev_embedding.expand(self.batch_size, -1, -1).clone() class RowState: # Per-row state tracking during generation def __init__(self, current_tokens=None): self.current_tokens = current_tokens or [] # Current token sequence for this row self.forced_tokens = deque() # Queue of tokens to force inject self.in_python_block = False # Whether we are inside a python block self.python_expr_tokens = [] # Tokens of the current python expression self.completed = False # Whether this row has completed generation @torch.inference_mode() def sample_next_token(logits, rng, temperature=1.0, top_k=None): """Sample a single next token from given logits of shape (B, vocab_size). Returns (B, 1).""" assert temperature >= 0.0, "temperature must be non-negative" if temperature == 0.0: return torch.argmax(logits, dim=-1, keepdim=True) if top_k is not None and top_k > 0: k = min(top_k, logits.size(-1)) vals, idx = torch.topk(logits, k, dim=-1) vals = vals / temperature probs = F.softmax(vals, dim=-1) choice = torch.multinomial(probs, num_samples=1, generator=rng) return idx.gather(1, choice) else: logits = logits / temperature probs = F.softmax(logits, dim=-1) return torch.multinomial(probs, num_samples=1, generator=rng) class Engine: def __init__(self, model, tokenizer): self.model = model self.tokenizer = tokenizer # needed for tool use @torch.inference_mode() def generate(self, tokens, negative_tokens=[], num_samples=1, max_tokens=None, temperature=1.0, top_k=None, seed=42): """Same as generate, but does single prefill and then clones the KV cache.""" assert isinstance(tokens, list) and isinstance(tokens[0], int), "expecting list of ints" device = self.model.get_device() # Allocate the KV cache in the compute dtype so it matches what the forward pass emits dtype = COMPUTE_DTYPE rng = torch.Generator(device=device) rng.manual_seed(seed) assistant_end = 1 # 1) Run a batch 1 prefill of the prompt tokens m = self.model.config kv_model_kwargs = {"num_heads": m.n_kv_head, "head_dim": m.n_embd // m.n_head, "num_layers": m.n_layer} kv_cache_prefill = KVCache( batch_size=1, seq_len=len(tokens), device=device, dtype=dtype, **kv_model_kwargs, ) ids = torch.tensor([tokens], dtype=torch.long, device=device) logits = self.model.forward(ids, kv_cache=kv_cache_prefill) logits = logits[:, -1, :].expand(num_samples, -1) # (num_samples, vocab_size) # 2) Replicate the KV cache for each sample/row kv_length_hint = (len(tokens) + max_tokens) if max_tokens is not None else self.model.config.sequence_len kv_cache_decode = KVCache( batch_size=num_samples, seq_len=kv_length_hint, device=device, dtype=dtype, **kv_model_kwargs, ) kv_cache_decode.prefill(kv_cache_prefill) del kv_cache_prefill # no need to keep this memory around # 3) Initialize states for each sample row_states = [RowState(tokens.copy()) for _ in range(num_samples)] # 4) Main generation loop num_generated = 0 while True: # Stop condition: we've reached max tokens if max_tokens is not None and num_generated >= max_tokens: break # Stop condition: all rows are completed if all(state.completed for state in row_states): break # Ban already-generated tags for each row for i, state in enumerate(row_states): banned = torch.tensor( list(set(state.current_tokens) | set(negative_tokens) - {0, 1}),#), # PAD=0, EOS=1 remain allowed dtype=torch.long, device=device, ) if len(banned) > 0: logits[i, banned] = -float("inf") # Sample the next token for each row next_ids = sample_next_token(logits, rng, temperature, top_k) # (B, 1) sampled_tokens = next_ids[:, 0].tolist() # Process each row: choose the next token, update state, optional tool use token_column = [] # contains the next token id along each row token_masks = [] # contains the mask (was it sampled (1) or forced (0)?) along each row for i, state in enumerate(row_states): # Select the next token in this row is_forced = len(state.forced_tokens) > 0 # are there tokens waiting to be forced in deque? token_masks.append(0 if is_forced else 1) # mask is 0 if forced, 1 if sampled next_token = state.forced_tokens.popleft() if is_forced else sampled_tokens[i] token_column.append(next_token) # Update the state of this row to include the next token state.current_tokens.append(next_token) # On <|assistant_end|> or <|bos|>, mark the row as completed if next_token == assistant_end: state.completed = True # Yield the token column yield token_column, token_masks num_generated += 1 # Prepare logits for next iteration ids = torch.tensor(token_column, dtype=torch.long, device=device).unsqueeze(1) logits = self.model.forward(ids, kv_cache=kv_cache_decode)[:, -1, :] # (B, vocab_size)