File size: 16,103 Bytes
c32c359 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 | """Minimal Qwen2 forward pass that consumes a paged KV cache.
We deliberately re-implement Qwen2 from scratch (rather than using the HF
forward) so the path of K/V tensors through the cache is fully visible.
Weights are loaded from a HuggingFace checkpoint by matching parameter names.
Layout of inputs per step ("varlen" packing):
input_ids [T_total] concatenated tokens for all seqs
positions [T_total] position-in-sequence of each token
slot_mapping [T_total] where to write new K/V in the cache
segments list of (q_start, q_end, block_table, k_len, seq_id)
For attention, we loop over `segments`: gather each sequence's full K/V from
its block table, run SDPA, scatter the result back into a flat buffer. All
other ops (norms, MLP, projections) run on the full packed tensor.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from .config import EngineConfig
from .paged_kv import PagedKVCache
from .request import Sequence
# ---------------------------------------------------------------------------
# Qwen2 building blocks
# ---------------------------------------------------------------------------
class Qwen2RMSNorm(nn.Module):
def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.eps = eps
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x: [..., hidden]
dtype = x.dtype
x = x.to(torch.float32)
var = x.pow(2).mean(-1, keepdim=True)
x = x * torch.rsqrt(var + self.eps)
return (self.weight * x).to(dtype)
def _rotate_half(x: torch.Tensor) -> torch.Tensor:
half = x.size(-1) // 2
x1, x2 = x[..., :half], x[..., half:]
return torch.cat((-x2, x1), dim=-1)
def _apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
"""x: [T, H, D], cos/sin: [T, D] → returns [T, H, D]."""
cos = cos.unsqueeze(1)
sin = sin.unsqueeze(1)
return (x * cos) + (_rotate_half(x) * sin)
class Qwen2MLP(nn.Module):
def __init__(self, hidden_size: int, intermediate_size: int) -> None:
super().__init__()
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
@dataclass
class AttnSegment:
"""One sequence's slice of the packed batch."""
q_start: int # start index in the packed tensor
q_end: int # exclusive
block_table: list[int] # KV blocks for this sequence
k_len: int # total K length (= num_computed_tokens + q_len)
class Qwen2Attention(nn.Module):
def __init__(
self,
hidden_size: int,
num_heads: int,
num_kv_heads: int,
head_dim: int,
layer_idx: int,
) -> None:
super().__init__()
self.num_heads = num_heads
self.num_kv_heads = num_kv_heads
self.head_dim = head_dim
self.layer_idx = layer_idx
self.q_proj = nn.Linear(hidden_size, num_heads * head_dim, bias=True)
self.k_proj = nn.Linear(hidden_size, num_kv_heads * head_dim, bias=True)
self.v_proj = nn.Linear(hidden_size, num_kv_heads * head_dim, bias=True)
self.o_proj = nn.Linear(num_heads * head_dim, hidden_size, bias=False)
self.scale = head_dim ** -0.5
def forward(
self,
hidden_states: torch.Tensor, # [T, hidden]
positions: torch.Tensor, # [T] long
slot_mapping: torch.Tensor, # [T] long
cos_table: torch.Tensor, # [max_pos, head_dim]
sin_table: torch.Tensor, # [max_pos, head_dim]
segments: list[AttnSegment],
kv_cache: PagedKVCache,
) -> torch.Tensor:
T = hidden_states.size(0)
q = self.q_proj(hidden_states).view(T, self.num_heads, self.head_dim)
k = self.k_proj(hidden_states).view(T, self.num_kv_heads, self.head_dim)
v = self.v_proj(hidden_states).view(T, self.num_kv_heads, self.head_dim)
cos = cos_table.index_select(0, positions) # [T, head_dim]
sin = sin_table.index_select(0, positions)
q = _apply_rope(q, cos, sin)
k = _apply_rope(k, cos, sin)
# Write the NEW K/V into the paged cache before reading it back.
kv_cache.write(self.layer_idx, k, v, slot_mapping)
out = torch.empty_like(q) # [T, num_heads, head_dim]
rep = self.num_heads // self.num_kv_heads # GQA fan-out
for seg in segments:
q_slice = q[seg.q_start:seg.q_end] # [q_len, H_q, D]
k_full, v_full = kv_cache.gather(self.layer_idx, seg.block_table, seg.k_len)
# GQA: expand K/V heads to match Q heads.
if rep > 1:
k_full = k_full.repeat_interleave(rep, dim=1)
v_full = v_full.repeat_interleave(rep, dim=1)
q_len = q_slice.size(0)
k_len = seg.k_len
num_past = k_len - q_len
# Causal mask: Q at logical position (num_past + i) attends to K at
# positions [0, num_past + i]. True = participate (SDPA convention).
idx_q = torch.arange(q_len, device=q.device).unsqueeze(1) + num_past
idx_k = torch.arange(k_len, device=q.device).unsqueeze(0)
attn_mask = idx_k <= idx_q # [q_len, k_len]
# SDPA wants [..., heads, q_len, head_dim]. Reshape and run.
q_h = q_slice.transpose(0, 1).unsqueeze(0) # [1, H, q_len, D]
k_h = k_full.transpose(0, 1).unsqueeze(0) # [1, H, k_len, D]
v_h = v_full.transpose(0, 1).unsqueeze(0)
attn = F.scaled_dot_product_attention(
q_h, k_h, v_h,
attn_mask=attn_mask.unsqueeze(0).unsqueeze(0), # [1,1,q_len,k_len]
scale=self.scale,
) # [1, H, q_len, D]
out[seg.q_start:seg.q_end] = attn.squeeze(0).transpose(0, 1)
return self.o_proj(out.reshape(T, self.num_heads * self.head_dim))
class Qwen2DecoderLayer(nn.Module):
def __init__(self, cfg: dict, layer_idx: int) -> None:
super().__init__()
self.input_layernorm = Qwen2RMSNorm(cfg["hidden_size"], eps=cfg["rms_norm_eps"])
self.self_attn = Qwen2Attention(
hidden_size=cfg["hidden_size"],
num_heads=cfg["num_attention_heads"],
num_kv_heads=cfg["num_key_value_heads"],
head_dim=cfg["head_dim"],
layer_idx=layer_idx,
)
self.post_attention_layernorm = Qwen2RMSNorm(cfg["hidden_size"], eps=cfg["rms_norm_eps"])
self.mlp = Qwen2MLP(cfg["hidden_size"], cfg["intermediate_size"])
def forward(self, hidden_states, positions, slot_mapping, cos_table, sin_table, segments, kv_cache):
residual = hidden_states
h = self.input_layernorm(hidden_states)
h = self.self_attn(h, positions, slot_mapping, cos_table, sin_table, segments, kv_cache)
hidden_states = residual + h
residual = hidden_states
h = self.post_attention_layernorm(hidden_states)
h = self.mlp(h)
return residual + h
class Qwen2Model(nn.Module):
def __init__(self, cfg: dict) -> None:
super().__init__()
self.cfg = cfg
self.embed_tokens = nn.Embedding(cfg["vocab_size"], cfg["hidden_size"])
self.layers = nn.ModuleList(
[Qwen2DecoderLayer(cfg, i) for i in range(cfg["num_hidden_layers"])]
)
self.norm = Qwen2RMSNorm(cfg["hidden_size"], eps=cfg["rms_norm_eps"])
def forward(self, input_ids, positions, slot_mapping, cos_table, sin_table, segments, kv_cache):
h = self.embed_tokens(input_ids)
for layer in self.layers:
h = layer(h, positions, slot_mapping, cos_table, sin_table, segments, kv_cache)
return self.norm(h)
class Qwen2ForCausalLM(nn.Module):
def __init__(self, cfg: dict) -> None:
super().__init__()
self.model = Qwen2Model(cfg)
self.lm_head = nn.Linear(cfg["hidden_size"], cfg["vocab_size"], bias=False)
self.cfg = cfg
def tie_weights(self) -> None:
self.lm_head.weight = self.model.embed_tokens.weight
# ---------------------------------------------------------------------------
# ModelRunner: prepares inputs, runs forward, extracts last-token logits.
# ---------------------------------------------------------------------------
@dataclass
class ModelInput:
input_ids: torch.Tensor
positions: torch.Tensor
slot_mapping: torch.Tensor
segments: list[AttnSegment]
# Index in the packed batch of the LAST token of each scheduled seq —
# that's where we'll read logits from for sampling.
last_token_indices: torch.Tensor
class ModelRunner:
def __init__(self, config: EngineConfig) -> None:
from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM
self.config = config
self.device = torch.device(config.device)
self.dtype = {
"float32": torch.float32,
"float16": torch.float16,
"bfloat16": torch.bfloat16,
}[config.dtype]
hf_cfg = AutoConfig.from_pretrained(
config.model, trust_remote_code=config.trust_remote_code
)
self.tokenizer = AutoTokenizer.from_pretrained(
config.model, trust_remote_code=config.trust_remote_code
)
model_type = getattr(hf_cfg, "model_type", "?")
if model_type not in ("qwen2", "qwen2_moe", "llama"):
# Llama-style works too because the math is identical; we issue a
# warning rather than a hard fail.
print(f"[tiny_vllm] WARNING: model_type={model_type!r}; expected qwen2-like. "
"Continuing — assuming Llama-compatible config.")
head_dim = getattr(hf_cfg, "head_dim", hf_cfg.hidden_size // hf_cfg.num_attention_heads)
cfg = {
"vocab_size": hf_cfg.vocab_size,
"hidden_size": hf_cfg.hidden_size,
"intermediate_size": hf_cfg.intermediate_size,
"num_hidden_layers": hf_cfg.num_hidden_layers,
"num_attention_heads": hf_cfg.num_attention_heads,
"num_key_value_heads": getattr(hf_cfg, "num_key_value_heads",
hf_cfg.num_attention_heads),
"head_dim": head_dim,
"rms_norm_eps": getattr(hf_cfg, "rms_norm_eps", 1e-6),
"rope_theta": getattr(hf_cfg, "rope_theta", 10000.0),
"max_position_embeddings": getattr(hf_cfg, "max_position_embeddings", 4096),
"tie_word_embeddings": getattr(hf_cfg, "tie_word_embeddings", False),
}
self.model_cfg = cfg
# Build our own model, then copy HF weights into it.
model = Qwen2ForCausalLM(cfg).to(self.device, self.dtype)
hf_model = AutoModelForCausalLM.from_pretrained(
config.model, torch_dtype=self.dtype,
trust_remote_code=config.trust_remote_code,
)
missing, unexpected = model.load_state_dict(hf_model.state_dict(), strict=False)
if cfg["tie_word_embeddings"] and "lm_head.weight" in (missing or []):
model.tie_weights()
del hf_model
model.eval()
for p in model.parameters():
p.requires_grad_(False)
self.model = model
# Precompute RoPE tables.
max_pos = min(cfg["max_position_embeddings"], config.max_model_len)
inv_freq = 1.0 / (
cfg["rope_theta"]
** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim)
)
t = torch.arange(max_pos, dtype=torch.float32)
freqs = torch.outer(t, inv_freq) # [max_pos, head_dim/2]
emb = torch.cat((freqs, freqs), dim=-1) # [max_pos, head_dim]
self.cos_table = emb.cos().to(self.device, self.dtype)
self.sin_table = emb.sin().to(self.device, self.dtype)
# Paged KV cache pool.
self.kv_cache = PagedKVCache(
num_layers=cfg["num_hidden_layers"],
num_blocks=config.num_blocks,
block_size=config.block_size,
num_kv_heads=cfg["num_key_value_heads"],
head_dim=head_dim,
dtype=self.dtype,
device=self.device,
)
# ---- input building ------------------------------------------------
def prepare_input(self, scheduled) -> ModelInput:
"""`scheduled` is a list of (Sequence, num_tokens, is_prefill) triples
from the scheduler."""
input_ids: list[int] = []
positions: list[int] = []
slot_mapping: list[int] = []
segments: list[AttnSegment] = []
last_indices: list[int] = []
cursor = 0
B = self.config.block_size
for item in scheduled:
seq = item.seq
n = item.num_tokens
# Logical token positions this step processes.
start_pos = seq.num_computed_tokens
for off in range(n):
pos = start_pos + off
input_ids.append(seq.get_token(pos))
positions.append(pos)
block_id = seq.block_table[pos // B]
slot_mapping.append(block_id * B + (pos % B))
q_end = cursor + n
segments.append(AttnSegment(
q_start=cursor,
q_end=q_end,
block_table=list(seq.block_table),
k_len=start_pos + n,
))
last_indices.append(q_end - 1)
cursor = q_end
return ModelInput(
input_ids=torch.tensor(input_ids, dtype=torch.long, device=self.device),
positions=torch.tensor(positions, dtype=torch.long, device=self.device),
slot_mapping=torch.tensor(slot_mapping, dtype=torch.long, device=self.device),
segments=segments,
last_token_indices=torch.tensor(last_indices, dtype=torch.long, device=self.device),
)
# ---- forward -------------------------------------------------------
@torch.inference_mode()
def execute(self, model_input: ModelInput) -> torch.Tensor:
"""Run one forward pass. Returns logits for the LAST token of each
scheduled sequence: shape [num_seqs, vocab_size]."""
hidden = self.model.model(
input_ids=model_input.input_ids,
positions=model_input.positions,
slot_mapping=model_input.slot_mapping,
cos_table=self.cos_table,
sin_table=self.sin_table,
segments=model_input.segments,
kv_cache=self.kv_cache,
) # [T, hidden]
last_hidden = hidden.index_select(0, model_input.last_token_indices)
logits = self.model.lm_head(last_hidden) # [num_seqs, vocab]
return logits
# ---- helpers -------------------------------------------------------
@property
def eos_token_id(self) -> Optional[int]:
return self.tokenizer.eos_token_id
def encode(self, text: str) -> list[int]:
return self.tokenizer.encode(text, add_special_tokens=False)
def decode(self, token_ids: list[int]) -> str:
return self.tokenizer.decode(token_ids, skip_special_tokens=True)
def detokenize_incremental(self, full_ids: list[int], prev_text_len: int) -> tuple[str, int]:
"""Detokenize the full list, return the new text added since last call
and the new total length."""
text = self.tokenizer.decode(full_ids, skip_special_tokens=True)
return text[prev_text_len:], len(text)
|