phase 9: KV cache + autoregressive generation
- KVCache: per-layer, per-head storage with append + reconstruct - forward_with_cache: prefill (full prompt) + decode (single token) modes - Fixed data layout bug: per-head vectors avoid cross-head interleaving - CLI updated to use KV cache by default - bench-gpt2 supports --no-cache flag for comparison Benchmark results (50 prompts × 20 tokens): - KV cache vs no-cache: 50/50 bit-identical (cache is correct) - 18x speedup: TTFT 400→24ms, TBT 407→22ms, throughput 2.5→44 tok/s - vs HF transformers: 40/50 match (10 are FP divergence, avg logit gap 0.20) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -6,27 +6,83 @@ use crate::config::ModelConfig;
|
||||
|
||||
pub struct GPT2 {
|
||||
pub config: ModelConfig,
|
||||
wte: Tensor, // [vocab_size, hidden]
|
||||
wpe: Tensor, // [max_pos, hidden]
|
||||
wte: Tensor,
|
||||
wpe: Tensor,
|
||||
layers: Vec<GPT2Block>,
|
||||
ln_f_g: Tensor, // [hidden]
|
||||
ln_f_b: Tensor, // [hidden]
|
||||
ln_f_g: Tensor,
|
||||
ln_f_b: Tensor,
|
||||
lm_head: Tensor, // precomputed wte^T
|
||||
}
|
||||
|
||||
struct GPT2Block {
|
||||
ln_1_g: Tensor,
|
||||
ln_1_b: Tensor,
|
||||
// Attention: combined QKV weight + bias, output weight + bias
|
||||
attn_qkv_w: Tensor, // [hidden, 3*hidden]
|
||||
attn_qkv_b: Tensor, // [3*hidden]
|
||||
attn_out_w: Tensor, // [hidden, hidden]
|
||||
attn_out_b: Tensor, // [hidden]
|
||||
attn_qkv_w: Tensor,
|
||||
attn_qkv_b: Tensor,
|
||||
attn_out_w: Tensor,
|
||||
attn_out_b: Tensor,
|
||||
ln_2_g: Tensor,
|
||||
ln_2_b: Tensor,
|
||||
mlp_fc_w: Tensor, // [hidden, 4*hidden]
|
||||
mlp_fc_b: Tensor, // [4*hidden]
|
||||
mlp_proj_w: Tensor, // [4*hidden, hidden]
|
||||
mlp_proj_b: Tensor, // [hidden]
|
||||
mlp_fc_w: Tensor,
|
||||
mlp_fc_b: Tensor,
|
||||
mlp_proj_w: Tensor,
|
||||
mlp_proj_b: Tensor,
|
||||
}
|
||||
|
||||
pub struct KVCache {
|
||||
// Per layer, per head: k[layer][head] has seq_len * head_dim floats
|
||||
k: Vec<Vec<Vec<f32>>>, // [num_layers][num_heads][seq_len * head_dim]
|
||||
v: Vec<Vec<Vec<f32>>>,
|
||||
len: usize,
|
||||
num_heads: usize,
|
||||
head_dim: usize,
|
||||
device: Device,
|
||||
}
|
||||
|
||||
impl KVCache {
|
||||
pub fn new(num_layers: usize, num_heads: usize, head_dim: usize, device: Device) -> Self {
|
||||
Self {
|
||||
k: (0..num_layers).map(|_| vec![vec![]; num_heads]).collect(),
|
||||
v: (0..num_layers).map(|_| vec![vec![]; num_heads]).collect(),
|
||||
len: 0,
|
||||
num_heads,
|
||||
head_dim,
|
||||
device,
|
||||
}
|
||||
}
|
||||
|
||||
pub fn seq_len(&self) -> usize { self.len }
|
||||
|
||||
/// Append new K/V data. k_new is in [1, H, new_tokens, D] layout (flat).
|
||||
fn append_kv(&mut self, layer: usize, k_new: &[f32], v_new: &[f32], new_tokens: usize) {
|
||||
let hd = self.head_dim;
|
||||
for h in 0..self.num_heads {
|
||||
let off = h * new_tokens * hd;
|
||||
self.k[layer][h].extend_from_slice(&k_new[off..off + new_tokens * hd]);
|
||||
self.v[layer][h].extend_from_slice(&v_new[off..off + new_tokens * hd]);
|
||||
}
|
||||
if layer == 0 {
|
||||
self.len += new_tokens;
|
||||
}
|
||||
}
|
||||
|
||||
/// Reconstruct [1, H, seq_len, D] tensors from per-head cache.
|
||||
fn get_kv_tensors(&self, layer: usize) -> (Tensor, Tensor) {
|
||||
let sl = self.len;
|
||||
let hd = self.head_dim;
|
||||
let nh = self.num_heads;
|
||||
let mut k_data = vec![0.0f32; nh * sl * hd];
|
||||
let mut v_data = vec![0.0f32; nh * sl * hd];
|
||||
for h in 0..nh {
|
||||
let off = h * sl * hd;
|
||||
k_data[off..off + sl * hd].copy_from_slice(&self.k[layer][h]);
|
||||
v_data[off..off + sl * hd].copy_from_slice(&self.v[layer][h]);
|
||||
}
|
||||
let shape = &[1, nh, sl, hd];
|
||||
let k = Tensor::from_slice(&k_data, shape).to_device(self.device);
|
||||
let v = Tensor::from_slice(&v_data, shape).to_device(self.device);
|
||||
(k, v)
|
||||
}
|
||||
}
|
||||
|
||||
impl GPT2 {
|
||||
@@ -39,6 +95,7 @@ impl GPT2 {
|
||||
let wpe = take(&mut w, "wpe.weight");
|
||||
let ln_f_g = take(&mut w, "ln_f.weight");
|
||||
let ln_f_b = take(&mut w, "ln_f.bias");
|
||||
let lm_head = wte.transpose(0, 1).contiguous();
|
||||
|
||||
let num_layers = config.num_layers();
|
||||
let mut layers = Vec::with_capacity(num_layers);
|
||||
@@ -60,81 +117,108 @@ impl GPT2 {
|
||||
});
|
||||
}
|
||||
|
||||
Self { config, wte, wpe, layers, ln_f_g, ln_f_b }
|
||||
Self { config, wte, wpe, layers, ln_f_g, ln_f_b, lm_head }
|
||||
}
|
||||
|
||||
/// Full forward pass, returns logits [seq_len, vocab_size].
|
||||
/// Full forward pass without KV cache (for testing / correctness comparison).
|
||||
pub fn forward(&self, token_ids: &[u32]) -> Tensor {
|
||||
let seq_len = token_ids.len();
|
||||
let hidden = self.config.hidden();
|
||||
let num_heads = self.config.num_heads();
|
||||
let head_dim = self.config.head_dim();
|
||||
|
||||
// Token + position embedding
|
||||
let tok_emb = embedding(&self.wte, token_ids);
|
||||
let pos_ids: Vec<u32> = (0..seq_len as u32).collect();
|
||||
let pos_emb = embedding(&self.wpe, &pos_ids);
|
||||
let mut x = add_tensors(&tok_emb, &pos_emb);
|
||||
|
||||
// Transformer layers
|
||||
for layer in &self.layers {
|
||||
// Pre-LN attention
|
||||
let residual = x.clone();
|
||||
let normed = layernorm(&x, &layer.ln_1_g, &layer.ln_1_b, self.config.ln_eps());
|
||||
|
||||
// QKV projection: [S, H] @ [H, 3H] + [3H] → [S, 3H]
|
||||
let qkv = linear(&normed, &layer.attn_qkv_w, Some(&layer.attn_qkv_b));
|
||||
// Split into Q, K, V and reshape for multi-head
|
||||
let (q, k, v) = split_qkv(&qkv, num_heads, head_dim, seq_len);
|
||||
// Attention: [1, H, S, D]
|
||||
let attn_out = attention(&q, &k, &v, true);
|
||||
// Merge heads: [1, H, S, D] → [S, hidden]
|
||||
let attn_out = merge_heads(&attn_out, seq_len, hidden);
|
||||
// Output projection
|
||||
let attn_out = linear(&attn_out, &layer.attn_out_w, Some(&layer.attn_out_b));
|
||||
x = add_tensors(&residual, &attn_out);
|
||||
|
||||
// Pre-LN MLP
|
||||
let residual = x.clone();
|
||||
let normed = layernorm(&x, &layer.ln_2_g, &layer.ln_2_b, self.config.ln_eps());
|
||||
let fc = linear(&normed, &layer.mlp_fc_w, Some(&layer.mlp_fc_b));
|
||||
let activated = gelu(&fc);
|
||||
let proj = linear(&activated, &layer.mlp_proj_w, Some(&layer.mlp_proj_b));
|
||||
x = add_tensors(&residual, &proj);
|
||||
x = self.transformer_block(layer, &x, None, 0, seq_len, num_heads, head_dim, hidden);
|
||||
}
|
||||
|
||||
// Final layer norm
|
||||
let x = layernorm(&x, &self.ln_f_g, &self.ln_f_b, self.config.ln_eps());
|
||||
matmul_2d(&x, &self.lm_head)
|
||||
}
|
||||
|
||||
// LM head (tied with wte): [S, H] @ [H, V] → [S, V]
|
||||
// wte is [V, H], so we need wte^T
|
||||
let lm_head = self.wte.transpose(0, 1).contiguous();
|
||||
matmul_2d(&x, &lm_head)
|
||||
/// Forward pass with KV cache. First call = prefill, subsequent = decode.
|
||||
pub fn forward_with_cache(&self, token_ids: &[u32], cache: &mut KVCache) -> Tensor {
|
||||
let new_tokens = token_ids.len();
|
||||
let pos_offset = cache.seq_len();
|
||||
let hidden = self.config.hidden();
|
||||
let num_heads = self.config.num_heads();
|
||||
let head_dim = self.config.head_dim();
|
||||
|
||||
let tok_emb = embedding(&self.wte, token_ids);
|
||||
let pos_ids: Vec<u32> = (pos_offset..pos_offset + new_tokens).map(|p| p as u32).collect();
|
||||
let pos_emb = embedding(&self.wpe, &pos_ids);
|
||||
let mut x = add_tensors(&tok_emb, &pos_emb);
|
||||
|
||||
for (layer_idx, layer) in self.layers.iter().enumerate() {
|
||||
x = self.transformer_block(
|
||||
layer, &x, Some((cache, layer_idx)),
|
||||
pos_offset, new_tokens, num_heads, head_dim, hidden,
|
||||
);
|
||||
}
|
||||
|
||||
let x = layernorm(&x, &self.ln_f_g, &self.ln_f_b, self.config.ln_eps());
|
||||
matmul_2d(&x, &self.lm_head)
|
||||
}
|
||||
|
||||
fn transformer_block(
|
||||
&self,
|
||||
layer: &GPT2Block,
|
||||
x: &Tensor,
|
||||
cache: Option<(&mut KVCache, usize)>,
|
||||
pos_offset: usize,
|
||||
new_tokens: usize,
|
||||
num_heads: usize,
|
||||
head_dim: usize,
|
||||
hidden: usize,
|
||||
) -> Tensor {
|
||||
let residual = x.clone();
|
||||
let normed = layernorm(x, &layer.ln_1_g, &layer.ln_1_b, self.config.ln_eps());
|
||||
|
||||
let qkv = linear(&normed, &layer.attn_qkv_w, Some(&layer.attn_qkv_b));
|
||||
let (q, k_new, v_new) = split_qkv(&qkv, num_heads, head_dim, new_tokens);
|
||||
|
||||
// KV cache: append new K/V, use full cached K/V for attention
|
||||
let (k_full, v_full) = if let Some((cache, layer_idx)) = cache {
|
||||
let k_cpu = k_new.to_device(Device::Cpu);
|
||||
let v_cpu = v_new.to_device(Device::Cpu);
|
||||
cache.append_kv(layer_idx, k_cpu.as_slice::<f32>(), v_cpu.as_slice::<f32>(), new_tokens);
|
||||
cache.get_kv_tensors(layer_idx)
|
||||
} else {
|
||||
(k_new, v_new)
|
||||
};
|
||||
|
||||
let attn_out = attention(&q, &k_full, &v_full, true);
|
||||
let attn_out = merge_heads(&attn_out, new_tokens, hidden);
|
||||
let attn_out = linear(&attn_out, &layer.attn_out_w, Some(&layer.attn_out_b));
|
||||
let x = add_tensors(&residual, &attn_out);
|
||||
|
||||
let residual = x.clone();
|
||||
let normed = layernorm(&x, &layer.ln_2_g, &layer.ln_2_b, self.config.ln_eps());
|
||||
let fc = linear(&normed, &layer.mlp_fc_w, Some(&layer.mlp_fc_b));
|
||||
let activated = gelu(&fc);
|
||||
let proj = linear(&activated, &layer.mlp_proj_w, Some(&layer.mlp_proj_b));
|
||||
add_tensors(&residual, &proj)
|
||||
}
|
||||
}
|
||||
|
||||
// --- Helper ops ---
|
||||
// --- Helper ops (unchanged) ---
|
||||
|
||||
fn linear(x: &Tensor, weight: &Tensor, bias: Option<&Tensor>) -> Tensor {
|
||||
// GPT-2 stores weights as [in, out] (not transposed), so x @ w
|
||||
let out = matmul_2d(x, weight);
|
||||
if let Some(b) = bias {
|
||||
add_bias(&out, b)
|
||||
} else {
|
||||
out
|
||||
}
|
||||
if let Some(b) = bias { add_bias(&out, b) } else { out }
|
||||
}
|
||||
|
||||
fn matmul_2d(a: &Tensor, b: &Tensor) -> Tensor {
|
||||
// a: [S, K], b: [K, N] → [S, N]
|
||||
assert_eq!(a.ndim(), 2);
|
||||
assert_eq!(b.ndim(), 2);
|
||||
matmul(a, b, GemmBackend::CuBlas)
|
||||
}
|
||||
|
||||
fn add_tensors(a: &Tensor, b: &Tensor) -> Tensor {
|
||||
// Element-wise add on GPU via a simple approach: scale(a, 1.0) + scale(b, 1.0)
|
||||
// TODO: proper add kernel. For now, go through CPU.
|
||||
assert_eq!(a.shape(), b.shape());
|
||||
assert_eq!(a.dtype(), DType::F32);
|
||||
let a_cpu = a.to_device(Device::Cpu);
|
||||
@@ -146,7 +230,6 @@ fn add_tensors(a: &Tensor, b: &Tensor) -> Tensor {
|
||||
}
|
||||
|
||||
fn add_bias(x: &Tensor, bias: &Tensor) -> Tensor {
|
||||
// x: [S, N], bias: [N] → broadcast add
|
||||
assert_eq!(x.ndim(), 2);
|
||||
assert_eq!(bias.ndim(), 1);
|
||||
assert_eq!(x.shape()[1], bias.shape()[0]);
|
||||
@@ -160,12 +243,10 @@ fn add_bias(x: &Tensor, bias: &Tensor) -> Tensor {
|
||||
}
|
||||
|
||||
fn split_qkv(qkv: &Tensor, num_heads: usize, head_dim: usize, seq_len: usize) -> (Tensor, Tensor, Tensor) {
|
||||
// qkv: [S, 3*H] → Q, K, V each [1, num_heads, S, head_dim]
|
||||
let hidden = num_heads * head_dim;
|
||||
let qkv_cpu = qkv.to_device(Device::Cpu);
|
||||
let data = qkv_cpu.as_slice::<f32>();
|
||||
|
||||
// Split into Q, K, V and directly write in [1, num_heads, S, head_dim] layout
|
||||
let mut q_data = vec![0.0f32; num_heads * seq_len * head_dim];
|
||||
let mut k_data = vec![0.0f32; num_heads * seq_len * head_dim];
|
||||
let mut v_data = vec![0.0f32; num_heads * seq_len * head_dim];
|
||||
@@ -189,14 +270,11 @@ fn split_qkv(qkv: &Tensor, num_heads: usize, head_dim: usize, seq_len: usize) ->
|
||||
}
|
||||
|
||||
fn merge_heads(x: &Tensor, seq_len: usize, hidden: usize) -> Tensor {
|
||||
// [1, num_heads, S, head_dim] → [S, hidden]
|
||||
let num_heads = x.shape()[1];
|
||||
let head_dim = x.shape()[3];
|
||||
let x_cpu = x.to_device(Device::Cpu);
|
||||
let src = x_cpu.as_slice::<f32>();
|
||||
|
||||
// src layout: [1][num_heads][seq_len][head_dim]
|
||||
// dst layout: [seq_len][hidden] where hidden = num_heads * head_dim
|
||||
let mut out = vec![0.0f32; seq_len * hidden];
|
||||
for s in 0..seq_len {
|
||||
for h in 0..num_heads {
|
||||
@@ -210,7 +288,7 @@ fn merge_heads(x: &Tensor, seq_len: usize, hidden: usize) -> Tensor {
|
||||
|
||||
/// Greedy sampling: return the argmax token ID from the last position's logits.
|
||||
pub fn sample_greedy(logits: &Tensor) -> u32 {
|
||||
assert_eq!(logits.ndim(), 2); // [S, V]
|
||||
assert_eq!(logits.ndim(), 2);
|
||||
let logits_cpu = logits.to_device(Device::Cpu);
|
||||
let data = logits_cpu.as_slice::<f32>();
|
||||
let vocab_size = logits.shape()[1];
|
||||
|
||||
Reference in New Issue
Block a user