happyz
/
gemma.cpp
mirror of https://github.com/google/gemma.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
gemma.cpp/gemma/gemma-inl.h

1634 lines
70 KiB
C++
Raw Blame History

// Copyright 2024 Google LLC
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// SIMD functions for Gemma/Griffin transformers.
#include <math.h> // sqrtf
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <algorithm> // std::min
#include <cstdio>
#include <vector>
#include "gemma/activations.h"
#include "gemma/common.h" // EmbeddingScaling
#include "gemma/configs.h"
#include "gemma/gemma.h"
#include "gemma/kv_cache.h"
#include "gemma/weights.h"
#include "paligemma/image.h"
#include "util/allocator.h"
#include "util/mat.h"
#include "util/threading_context.h"
#include "hwy/aligned_allocator.h" // Span
#include "hwy/contrib/thread_pool/thread_pool.h"
#include "hwy/profiler.h"
#include "hwy/timer.h"
// Include guard (still compiled once per target)
#if defined(THIRD_PARTY_GEMMA_CPP_GEMMA_GEMMA_INL_H_) == \
defined(HWY_TARGET_TOGGLE)
#ifdef THIRD_PARTY_GEMMA_CPP_GEMMA_GEMMA_INL_H_
#undef THIRD_PARTY_GEMMA_CPP_GEMMA_GEMMA_INL_H_
#else
#define THIRD_PARTY_GEMMA_CPP_GEMMA_GEMMA_INL_H_
#endif
#include "hwy/highway.h"
// After highway.h
#include "ops/matmul-inl.h"
#include "ops/matvec-inl.h"
#include "ops/ops-inl.h"
#ifndef GEMMA_TYPE
#if HWY_IDE
// Provide a definition so the IDE does not complain.
#define GEMMA_TYPE float
#else
#error "Only include from instantiations/*.cc, which must define GEMMA_TYPE"
#endif // HWY_IDE
#endif // GEMMA_TYPE
HWY_BEFORE_NAMESPACE();
namespace gcpp {
namespace HWY_NAMESPACE {
// Different functions use different naming conventions for the number of
// tokens. Functions that are query-independent, such as RMSNorm*, call the
// count `num_interleaved`. Functions that are query-dependent, such as
// `Attention`, use separate `num_tokens` and `num_queries`.
// TODO: add batch query support for Griffin (QueriesPos).
template <typename T>
HWY_NOINLINE void GriffinRecurrent(size_t batch_start, size_t num_tokens,
size_t layer, Activations& activations,
const LayerWeightsPtrs<T>* layer_weights,
const KVCaches& kv_caches) {
PROFILER_ZONE("Gen.Griffin");
KVCache& kv_cache = kv_caches[0];
hwy::ThreadPool& pool = activations.env->ctx.pools.Pool(0);
namespace hn = hwy::HWY_NAMESPACE;
using D = hn::ScalableTag<float>;
const size_t model_dim = layer_weights->layer_config.model_dim;
const size_t conv_1d_width = layer_weights->layer_config.conv1d_width;
const size_t heads = layer_weights->layer_config.heads;
// X / Y linear layers.
for (size_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
float* HWY_RESTRICT y = activations.griffin_y.Row(batch_idx);
float* HWY_RESTRICT x = activations.griffin_x.Row(batch_idx);
TwoMatVecAdd(layer_weights->griffin.linear_x_w,
layer_weights->griffin.linear_y_w, 0, model_dim, model_dim,
activations.pre_att_rms_out.Row(batch_idx),
/*add0=*/layer_weights->griffin.linear_x_biases.PackedScale1(),
/*add1=*/layer_weights->griffin.linear_y_biases.PackedScale1(),
/*out0=*/x, /*out1=*/y, pool);
Gelu(y, model_dim);
}
// Conv1D.
for (size_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
const size_t pos = batch_start + batch_idx;
float* HWY_RESTRICT x = activations.griffin_x.Row(batch_idx);
HWY_FULL(float) df;
HWY_DASSERT(model_dim % hn::Lanes(df) == 0);
// cache[i] = input at time t-i.
float* HWY_RESTRICT cache[kMaxConv1DWidth];
cache[0] = x;
for (size_t i = 1; i < conv_1d_width; i++) {
cache[i] =
kv_cache.conv1d_cache.Row(layer) +
((pos + conv_1d_width - 1 - i) % (conv_1d_width - 1)) * model_dim;
}
for (size_t i = 0; i < model_dim; i += hn::Lanes(df)) {
auto xv = hn::Load(df, x + i);
auto accum0 =
hn::Load(df, layer_weights->griffin.conv_biases.PackedScale1() + i);
auto accum1 = hn::Zero(df);
HWY_ASSERT_M(conv_1d_width % 2 == 0, "Conv width must be even");
for (size_t l = 0; 2 * l < conv_1d_width; l++) {
auto wv0 =
hn::Load(df, layer_weights->griffin.conv_w.PackedScale1() +
(conv_1d_width - 1 - 2 * l) * model_dim + i);
auto wv1 =
hn::Load(df, layer_weights->griffin.conv_w.PackedScale1() +
(conv_1d_width - 2 - 2 * l) * model_dim + i);
accum0 = hn::MulAdd(wv0, hn::Load(df, cache[l * 2] + i), accum0);
accum1 = hn::MulAdd(wv1, hn::Load(df, cache[l * 2 + 1] + i), accum1);
}
hn::Store(hn::Add(accum0, accum1), df, x + i);
hn::Store(xv, df, cache[HWY_MAX(conv_1d_width, 1) - 1] + i);
}
}
// RGLRU
for (size_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
const size_t pos = batch_start + batch_idx;
float* HWY_RESTRICT y = activations.griffin_y.Row(batch_idx);
float* HWY_RESTRICT x = activations.griffin_x.Row(batch_idx);
float* HWY_RESTRICT gate_x = activations.griffin_gate_x.Row(batch_idx);
float* HWY_RESTRICT a = activations.griffin_multiplier.Row(batch_idx);
float* HWY_RESTRICT rnn_state = kv_cache.rglru_cache.Row(layer);
pool.Run(0, heads, [&](const uint64_t head, size_t /*thread*/) HWY_ATTR {
const size_t kHeadDim = model_dim / heads;
const size_t kMatrixSize = kHeadDim * kHeadDim;
size_t head_offset = head * kHeadDim;
TwoOfsMatVecAddLoop(
layer_weights->griffin.gate_w, kMatrixSize * head,
kMatrixSize * (heads + head), kHeadDim, kHeadDim, x + head_offset,
/*add0=*/layer_weights->griffin.gate_biases.PackedScale1() +
head_offset,
/*add1=*/layer_weights->griffin.gate_biases.PackedScale1() +
model_dim + head_offset,
/*out0=*/gate_x + head_offset, /*out1=*/a + head_offset);
Sigmoid(gate_x + head_offset, kHeadDim);
Sigmoid(a + head_offset, kHeadDim);
const auto fn_mul = [](D d, hn::Vec<D> x, hn::Vec<D> gate_x)
HWY_ATTR { return hn::Mul(x, gate_x); };
hn::Transform1(D(), a + head_offset, kHeadDim,
layer_weights->griffin.a.PackedScale1() + head_offset,
fn_mul);
hn::Transform1(D(), x + head_offset, kHeadDim, gate_x + head_offset,
fn_mul);
// RNN scan
HWY_FULL(float) df;
HWY_DASSERT(kHeadDim % hn::Lanes(df) == 0);
for (size_t i = 0; i < kHeadDim; i += hn::Lanes(df)) {
auto log_a = hn::Load(df, a + head_offset + i);
auto gated_x = hn::Load(df, x + head_offset + i);
auto rnn = hn::Load(df, rnn_state + head_offset + i);
auto a = hn::Exp(df, log_a);
auto x_multiplier = hn::Sqrt(hn::NegMulAdd(a, a, hn::Set(df, 1.0f)));
if (pos == 0) {
x_multiplier = hn::Set(df, 1.0f);
}
auto new_x = hn::MulAdd(x_multiplier, gated_x, hn::Mul(a, rnn));
hn::Store(new_x, df, rnn_state + head_offset + i);
// Join branches.
auto yv = hn::Load(df, y + head_offset + i);
auto pre_out = hn::Mul(yv, new_x);
hn::Store(pre_out, df, x + head_offset + i);
}
});
}
// Final linear layer.
for (size_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
float* HWY_RESTRICT x = activations.griffin_x.Row(batch_idx);
float* out_ptr = activations.att_sums.Row(batch_idx);
MatVecAdd(layer_weights->griffin.linear_out_w, 0, model_dim, model_dim, x,
layer_weights->griffin.linear_out_biases.PackedScale1(), out_ptr,
pool);
}
}
// Wrapper class; holds arguments in member variables to shorten call sites.
template <typename T>
class GemmaAttention {
// The attention window usually starts at 0 unless `pos` is larger than
// the attention window size, then it is `pos` - window_size + 1.
HWY_INLINE size_t StartPos(size_t pos, size_t layer) {
const size_t att_window_size =
activations_.weights_config.attention_window_sizes[layer];
return pos - std::min(att_window_size - 1, pos);
}
template <typename U>
HWY_INLINE void PositionalEncodingQK(U* qk, size_t pos, size_t layer,
const float mul) {
// qk is either q or k, so qkv_dim is the length we operate on.
const size_t qkv_dim = layer_config_.qkv_dim;
const float* inv_timescale = activations_.inv_timescale.Packed();
bool is_global_layer =
activations_.weights_config.attention_window_sizes[layer] ==
activations_.seq_len;
// TODO: add a config flag instead of hardcoding the model.
if (is_global_layer &&
(activations_.weights_config.model == Model::GEMMA3_4B ||
activations_.weights_config.model == Model::GEMMA3_12B ||
activations_.weights_config.model == Model::GEMMA3_27B ||
activations_.weights_config.model == Model::GEMMA3_1B)) {
inv_timescale = activations_.inv_timescale_global.Packed();
}
// PostQKType::Rope
(void)layer;
if (layer_weights_.layer_config.post_qk == PostQKType::HalfRope) {
Rope(qk, qkv_dim / 2, inv_timescale, pos);
if (mul != 1.0f) MulByConst(mul, qk, qkv_dim);
} else {
RopeAndMulBy(mul, qk, qkv_dim, inv_timescale, pos);
}
}
// Fills activations.q and computes KV. For is_mha_, a single MatMul suffices
// and we later copy KV from q to KVCache. Otherwise, a second MatMul writes
// KV directly to KVCache.
HWY_NOINLINE void ComputeQKV(const size_t num_interleaved) {
PROFILER_ZONE("Gen.Attention.QKV");
const size_t model_dim = layer_config_.model_dim;
const size_t qkv_dim = layer_config_.qkv_dim;
const size_t heads = layer_config_.heads;
const size_t kv_heads = layer_config_.kv_heads;
using WeightT = typename decltype(layer_weights_.qkv_einsum_w)::T;
ConstMat<WeightT> w_q1(layer_weights_.qkv_einsum_w.HasPtr()
? layer_weights_.qkv_einsum_w
: layer_weights_.qkv_einsum_w1);
// The original qkv_einsum_w has shape [(heads + kv_heads * 2), kKQVDim,
// model_dim], which we reshaped to (heads + kv_heads * 2) * kKQVDim rows.
// We must shrink to the actual size because MatMul verifies
// `B.extents.rows == C.Cols()`. If MHA, `QStride() == 3 * qkv_dim` and all
// rows are used. Otherwise, `QStride() == qkv_dim` and KV will be
// computed in the second MatMul.
const size_t w1_rows = heads * layer_config_.QStride();
w_q1.ShrinkRows(w1_rows);
MatMul(activations_.pre_att_rms_out, w_q1,
/*add=*/nullptr, *activations_.env,
RowPtrFromMat(allocator_, activations_.q));
if (is_mha_) {
// Multi-Head Attention a.k.a. "use_qkv_einsum" computed QKV already.
} else {
decltype(w_q1) w_q2(layer_weights_.qkv_einsum_w.HasPtr()
? layer_weights_.qkv_einsum_w
: layer_weights_.qkv_einsum_w2);
if (layer_weights_.qkv_einsum_w.HasPtr()) {
// Skip first half of the matrix.
w_q2.ofs = w_q2.Row(w1_rows);
}
// KV structure is [k, v, k, v, ....] = kv_heads pairs of (k, v).
const size_t w_rows_kv_cols = kv_heads * 2 * qkv_dim;
w_q2.ShrinkRows(w_rows_kv_cols);
// Single query and no wraparound means we can use a matmul and write
// directly into the KV cache with a stride of cache_pos_size_.
if (num_queries_ == 1 &&
queries_pos_[0] + num_tokens_ <= div_seq_len_.GetDivisor()) {
const size_t kv_ofs =
queries_pos_[0] * cache_pos_size_ + layer_ * cache_layer_size_;
float* HWY_RESTRICT kv = kv_caches_[0].kv_cache.get() + kv_ofs;
RowPtrF kv_rows(allocator_, kv, w_rows_kv_cols);
kv_rows.SetStride(cache_pos_size_);
MatMul(activations_.pre_att_rms_out, w_q2,
/*add=*/nullptr, *activations_.env, kv_rows);
} else {
// Proceed row by row because there will be wraparound.
for (size_t interleaved_idx = 0; interleaved_idx < num_interleaved;
++interleaved_idx) {
const float* x = activations_.pre_att_rms_out.Row(interleaved_idx);
const size_t query_idx = interleaved_idx % num_queries_;
const size_t batch_idx = interleaved_idx / num_queries_;
KVCache& kv_cache = kv_caches_[query_idx];
const size_t cache_pos =
div_seq_len_.Remainder(queries_pos_[query_idx] + batch_idx);
const size_t kv_offset =
cache_pos * cache_pos_size_ + layer_ * cache_layer_size_;
float* HWY_RESTRICT kv = kv_cache.kv_cache.get() + kv_offset;
MatVec(w_q2, w_q2.ofs, w_rows_kv_cols, model_dim, x, kv, pool_);
}
}
} // !is_mha_
// Apply positional encodings for K (and copy KV to cache if MHA).
pool_.Run(0, kv_heads * num_interleaved,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % kv_heads;
const size_t interleaved_idx = task / kv_heads;
const size_t query_idx = interleaved_idx % num_queries_;
const size_t batch_idx = interleaved_idx / num_queries_;
const size_t pos = queries_pos_[query_idx] + batch_idx;
const size_t cache_pos = div_seq_len_.Remainder(pos);
const size_t kv_offset = cache_pos * cache_pos_size_ +
layer_ * cache_layer_size_ +
head * qkv_dim * 2;
KVCache& kv_cache = kv_caches_[query_idx];
float* HWY_RESTRICT kv = kv_cache.kv_cache.get() + kv_offset;
// If MHA, copy computed K and V into KVCache.
if (is_mha_) {
const float* HWY_RESTRICT mha_kv =
activations_.q.Row(interleaved_idx) + head * q_stride_ +
qkv_dim;
hwy::CopyBytes(mha_kv, kv, 2 * qkv_dim * sizeof(*kv));
}
// Apply further processing to K.
if (layer_weights_.key_norm_scale.HasPtr()) {
RMSNormInplace(layer_weights_.key_norm_scale.PackedScale1(),
0, kv, qkv_dim);
}
PositionalEncodingQK(kv, pos, layer_, /*mul=*/1.0f);
});
}
// Computes Q.K scores, which are "logits" (or scores) stored to att.
// `k` is a strided view of the kv cache with dimensions [seq_len, qkv_dim].
HWY_INLINE void QDotK(const size_t start_pos, const size_t last_pos,
const float* HWY_RESTRICT q, const ConstMat<float>& k,
float* HWY_RESTRICT att) {
const size_t qkv_dim = layer_config_.qkv_dim;
if (HWY_LIKELY(last_pos < activations_.seq_len)) {
// Slightly faster: no wraparound.
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const float* HWY_RESTRICT k_ptr = k.ptr + k.Row(pos);
const float score = Dot(q, k_ptr, qkv_dim);
att[pos] = score;
}
} else {
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const size_t cache_pos = div_seq_len_.Remainder(pos);
const float* HWY_RESTRICT k_ptr = k.ptr + k.Row(cache_pos);
const float score = Dot(q, k_ptr, qkv_dim);
att[pos % activations_.seq_len] = score;
}
}
}
// Accumulates the sum of v (from `kv_cache`) * probability (`att`) into
// `att_out`. Equivalent in gemma/modules.py:
// encoded = jnp.einsum('BTNS,BSNH->BTNH', probs, value_proj)
// `v` is a strided view of the kv cache with dimensions [seq_len, qkv_dim].
HWY_INLINE void WeightedSumV(const size_t start_pos, const size_t last_pos,
const float* HWY_RESTRICT att,
const ConstMat<float>& v,
float* HWY_RESTRICT att_out) const {
const size_t qkv_dim = layer_config_.qkv_dim;
hwy::ZeroBytes(att_out, qkv_dim * sizeof(*att_out));
if (HWY_LIKELY(last_pos < activations_.seq_len)) {
// Slightly faster: no wraparound.
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const float* HWY_RESTRICT v_ptr = v.ptr + v.Row(pos);
MulByConstAndAdd(att[pos], v_ptr, att_out, qkv_dim);
}
} else {
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const size_t cache_pos = div_seq_len_.Remainder(pos);
const float* HWY_RESTRICT v_ptr = v.ptr + v.Row(cache_pos);
MulByConstAndAdd(att[pos % activations_.seq_len], v_ptr, att_out,
qkv_dim);
}
}
}
public:
// Calculates the attention outputs for a single q.
HWY_INLINE void SingleDotSoftmaxWeightedSum(
float* HWY_RESTRICT q, const ConstMat<float>& k, const ConstMat<float>& v,
float* HWY_RESTRICT att, float* HWY_RESTRICT att_out,
const float query_scale, const size_t pos, const size_t start_pos,
const size_t last_pos) {
const size_t qkv_dim = layer_config_.qkv_dim;
// Apply rope and scaling to Q.
if (layer_weights_.query_norm_scale.HasPtr()) {
RMSNormInplace(layer_weights_.query_norm_scale.PackedScale1(), 0, q,
qkv_dim);
}
PositionalEncodingQK(q, pos, layer_, query_scale);
QDotK(start_pos, last_pos, q, k, att);
// SoftMax with optional SoftCap yields "probabilities" in att.
const size_t att_len = std::min(last_pos + 1, activations_.seq_len);
MaybeLogitsSoftCap(activations_.weights_config.att_cap, att, att_len);
Softmax(att, att_len);
WeightedSumV(start_pos, last_pos, att, v, att_out);
}
HWY_NOINLINE void DotSoftmaxWeightedSum(const size_t num_interleaved) {
PROFILER_ZONE("Gen.Attention.DotSoftmax");
const float query_scale = ChooseQueryScale(activations_.weights_config);
// A "head group" in the context of GQA refers to a collection of query
// heads that share the same key and value heads.
const size_t kHeadGroups = layer_config_.heads / layer_config_.kv_heads;
// For each head (token, query), compute Q.K, softmax, and weighted V.
pool_.Run(0, layer_config_.heads * num_interleaved,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % layer_config_.heads;
const size_t interleaved_idx = task / layer_config_.heads;
const size_t query_idx = interleaved_idx % num_queries_;
const size_t batch_idx = interleaved_idx / num_queries_;
const size_t qkv_dim = layer_config_.qkv_dim;
const size_t head_offset = (head / kHeadGroups) * qkv_dim * 2;
float* HWY_RESTRICT q =
activations_.q.Row(interleaved_idx) + head * q_stride_;
float* HWY_RESTRICT att =
activations_.att.Row(interleaved_idx) +
head * activations_.seq_len;
float* HWY_RESTRICT att_out =
activations_.att_out.Row(interleaved_idx) + head * qkv_dim;
// Make strided views into the kv cache entries for the current
// query and head.
KVCache& kv_cache = kv_caches_[query_idx];
const size_t kv_head_offset =
layer_ * cache_layer_size_ + head_offset;
ConstMat<float> k(kv_cache.kv_cache.get() + kv_head_offset,
Extents2D(kv_cache.seq_len, qkv_dim),
/*stride=*/cache_pos_size_);
ConstMat<float> v(
kv_cache.kv_cache.get() + kv_head_offset + qkv_dim,
Extents2D(kv_cache.seq_len, qkv_dim),
/*stride=*/cache_pos_size_);
// Find the token position in the query and calculate the range
// of cache positions to attend to.
const size_t pos = queries_pos_[query_idx] + batch_idx;
const size_t start_pos = StartPos(pos, layer_);
size_t last_pos = pos;
const size_t prefix_end = queries_prefix_end_[query_idx];
if (prefix_end > 0 && prefix_end - 1 > last_pos) {
// last_pos in QDotK and WeightedSumV is inclusive.
last_pos = prefix_end - 1;
}
SingleDotSoftmaxWeightedSum(q, k, v, att, att_out, query_scale,
pos, start_pos, last_pos);
});
}
private:
// Sums encoded (`att_out`) over num_heads (`layer_config_.heads`) and
// head_dim (`qkv_dim`) into output (`layer_out`).
HWY_NOINLINE void SumHeads() {
PROFILER_ZONE("Gen.Attention.SumHeads");
// att_weights and att_out are concatenated heads, each of length
// layer_config_.qkv_dim. Thus the [num_interleaved,
// layer_config_.model_dim] matmul output is the sum over heads. Compare
// gemma/modules.py: attn_output = self.attn_vec_einsum('BTNH,NHD->BTD',
// encoded)
HWY_DASSERT(layer_config_.model_dim != 0 && layer_config_.heads != 0 &&
layer_config_.qkv_dim != 0);
HWY_DASSERT(layer_weights_.att_weights.HasPtr());
HWY_DASSERT(activations_.att_out.HasPtr());
HWY_DASSERT(activations_.att_sums.HasPtr());
const float* add =
layer_weights_.layer_config.softmax_attn_output_biases
? layer_weights_.attention_output_biases.PackedScale1()
: nullptr;
MatMul(activations_.att_out, layer_weights_.att_weights, add,
*activations_.env, RowPtrFromMat(allocator_, activations_.att_sums));
}
public:
// Constructor with explicit initialization of queries_prefix_end. This is
// needed for the Prefix-LM style attention. For standard causal attention,
// the other constructor can be used.
GemmaAttention(const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end, size_t num_tokens,
size_t layer, Activations& activations,
const LayerWeightsPtrs<T>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches)
: GemmaAttention(queries_pos, &queries_prefix_end, num_tokens, layer,
activations, layer_weights, div_seq_len, kv_caches) {}
// Constructor with default initialization to 0 for queries_prefix_end.
GemmaAttention(const QueriesPos& queries_pos, size_t num_tokens, size_t layer,
Activations& activations,
const LayerWeightsPtrs<T>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches)
: GemmaAttention(queries_pos, nullptr, num_tokens, layer, activations,
layer_weights, div_seq_len, kv_caches) {}
// Full attention computation in three steps.
HWY_INLINE void operator()() {
const size_t num_interleaved = num_tokens_ * num_queries_;
ComputeQKV(num_interleaved);
DotSoftmaxWeightedSum(num_interleaved);
SumHeads();
}
private:
// Delegated Constructor that does most of the common work.
GemmaAttention(const QueriesPos& queries_pos,
const QueriesPos* queries_prefix_end, size_t num_tokens,
size_t layer, Activations& activations,
const LayerWeightsPtrs<T>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches)
: queries_pos_(queries_pos),
num_queries_(queries_pos.size()),
num_tokens_(num_tokens),
layer_(layer),
layer_config_(layer_weights->layer_config),
q_stride_(layer_config_.QStride()),
cache_layer_size_(layer_weights->layer_config.CacheLayerSize()),
cache_pos_size_(activations.cache_pos_size),
is_mha_(layer_config_.IsMHA()),
activations_(activations),
layer_weights_(*layer_weights),
div_seq_len_(div_seq_len),
kv_caches_(kv_caches),
allocator_(activations.env->ctx.allocator),
pool_(activations.env->ctx.pools.Pool(0)) {
HWY_DASSERT(num_queries_ <= kv_caches_.size());
HWY_DASSERT_M((layer_config_.heads % layer_config_.kv_heads) == 0,
"query heads must be a multiple of key-value heads");
if (queries_prefix_end != nullptr) {
queries_prefix_end_ = *queries_prefix_end;
} else {
queries_prefix_end_vec_.assign(num_queries_, 0);
queries_prefix_end_ = QueriesPos(queries_prefix_end_vec_.data(),
queries_prefix_end_vec_.size());
}
}
const QueriesPos& queries_pos_;
std::vector<size_t> queries_prefix_end_vec_;
QueriesPos queries_prefix_end_;
const size_t num_queries_;
const size_t num_tokens_;
const size_t layer_;
const LayerConfig& layer_config_;
const size_t q_stride_ = 0;
const size_t cache_layer_size_ = 0;
const size_t cache_pos_size_ = 0;
const bool is_mha_ = false;
Activations& activations_;
const LayerWeightsPtrs<T>& layer_weights_;
const hwy::Divisor& div_seq_len_;
const KVCaches& kv_caches_;
const Allocator& allocator_;
hwy::ThreadPool& pool_;
};
template <typename T>
HWY_NOINLINE void Attention(
LayerAttentionType type, const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end, size_t num_tokens, size_t layer,
Activations& activations, const LayerWeightsPtrs<T>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches) {
if (type == LayerAttentionType::kGemma) {
GemmaAttention<T>(queries_pos, queries_prefix_end, num_tokens, layer,
activations, layer_weights, div_seq_len, kv_caches)();
} else {
// Only reached if the model is Griffin.
// The kv_caches are allocated only for the griffin layers, so we need to
// map the layer index to the griffin layer index.
auto type = layer_weights->layer_config.type;
size_t layer_of_type =
activations.weights_config.NumLayersOfTypeBefore(type, layer);
HWY_ASSERT(queries_pos.size() == 1);
GriffinRecurrent(queries_pos[0], num_tokens, layer_of_type, activations,
layer_weights, kv_caches);
}
}
// Wrapper class; holds arguments in member variables to shorten call sites.
// The main differences to GemmaAttention are:
// - no KV Cache necessary, attention is always all-to-all and not causal.
// - no potential wrap-around, attention always goes from 0 to kSeqLen.
// - no need for batching, as we are always computing attention for kSeqLen
// tokens.
// This results in a much simpler implementation. However, to avoid duplicating
// code, we should still consider merging the two classes.
// TODO(keysers): Refactor to share code with GemmaAttention.
template <typename T>
class VitAttention {
// Computes Q, K, V for all heads, stored in activations_.q.
HWY_NOINLINE void ComputeQKV() {
PROFILER_ZONE("Gen.VitAttention.QKV");
auto& qkv = activations_.q;
HWY_ASSERT(qkv.Rows() == num_tokens_);
HWY_ASSERT(qkv.Cols() == layer_config_.heads * 3 * layer_config_.qkv_dim);
MatMul(activations_.pre_att_rms_out, layer_weights_.vit.qkv_einsum_w,
layer_weights_.vit.qkv_einsum_b.PackedScale1(), *activations_.env,
RowPtrFromMat(allocator_, qkv));
}
// TODO(philculliton): transition fully to MatMul.
HWY_NOINLINE void DotSoftmaxWeightedSumMatrix() {
const size_t qkv_dim = layer_config_.qkv_dim;
const size_t heads = layer_config_.heads;
HWY_ASSERT_M(heads == layer_config_.kv_heads, "Vit expects MHA");
const size_t seq_len = activations_.seq_len;
const float query_scale = 1.0f / sqrtf(static_cast<float>(qkv_dim));
PROFILER_ZONE("Gen.VitAttention.DotSoftmax");
// Shift Q, K, VT to MatStorageT.
MatStorageT<float> Q("Q2", Extents2D(num_tokens_, qkv_dim),
MatPadding::kPacked);
MatStorageT<float> K("K2", Extents2D(seq_len, qkv_dim),
MatPadding::kPacked);
MatStorageT<float> C("C2", Extents2D(num_tokens_, seq_len),
MatPadding::kPacked);
// Initialize att_out to zero prior to head loop.
ZeroInit(activations_.att_out);
for (size_t head = 0; head < heads; ++head) {
pool_.Run(0, num_tokens_, [&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t token = task;
float* HWY_RESTRICT q = activations_.q.Row(token) + head * 3 * qkv_dim;
// TODO: shift to MatMul with A.scale once MatMul is confirmed working
MulByConst(query_scale, q, qkv_dim);
hwy::CopyBytes(q, Q.Row(token), qkv_dim * sizeof(float));
});
pool_.Run(0, seq_len, [&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t seq_idx = task;
float* HWY_RESTRICT k =
activations_.q.Row(seq_idx) + head * 3 * qkv_dim + qkv_dim;
hwy::CopyBytes(k, K.Row(seq_idx), qkv_dim * sizeof(float));
});
// this produces C, a (num_tokens_, seq_len) matrix of dot products
MatMul(Q, K, nullptr, *activations_.env, RowPtrFromMat(allocator_, C));
pool_.Run(0, num_tokens_, [&](uint64_t task, size_t /*thread*/) HWY_ATTR {
float* HWY_RESTRICT c = C.Row(task);
Softmax(c, C.Cols());
});
pool_.Run(0, num_tokens_, [&](uint64_t task, size_t /*thread*/) HWY_ATTR {
size_t token = task;
float* HWY_RESTRICT att_out =
activations_.att_out.Row(token) + head * qkv_dim;
for (size_t i = 0; i < seq_len; ++i) {
float* HWY_RESTRICT v =
activations_.q.Row(i) + head * 3 * qkv_dim + 2 * qkv_dim;
MulByConstAndAdd(C.Row(token)[i], v, att_out, qkv_dim);
}
});
}
}
HWY_NOINLINE void DotSoftmaxWeightedSum() {
const size_t qkv_dim = layer_config_.qkv_dim;
const size_t heads = layer_config_.heads;
HWY_ASSERT_M(heads == layer_config_.kv_heads, "Vit expects MHA");
const size_t seq_len = activations_.seq_len;
const float query_scale = 1.0f / sqrtf(static_cast<float>(qkv_dim));
PROFILER_ZONE("Gen.VitAttention.DotSoftmax");
// Compute Q.K, softmax, and weighted V.
pool_.Run(0, layer_config_.heads * num_tokens_,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % layer_config_.heads;
const size_t token = task / layer_config_.heads;
// Compute Q.K scores, which are "logits" stored in head_att.
float* HWY_RESTRICT q =
activations_.q.Row(token) + head * 3 * qkv_dim;
MulByConst(query_scale, q, qkv_dim);
float* HWY_RESTRICT head_att =
activations_.att.Row(token) + head * activations_.seq_len;
for (size_t i = 0; i < seq_len; ++i) {
float* HWY_RESTRICT k =
activations_.q.Row(i) + head * 3 * qkv_dim + qkv_dim;
head_att[i] = Dot(q, k, qkv_dim); // score = q.k
}
// SoftMax yields "probabilities" in head_att.
Softmax(head_att, seq_len);
// Compute weighted sum of v into att_out.
float* HWY_RESTRICT att_out =
activations_.att_out.Row(token) + head * qkv_dim;
hwy::ZeroBytes(att_out, qkv_dim * sizeof(*att_out));
for (size_t i = 0; i < seq_len; ++i) {
float* HWY_RESTRICT v =
activations_.q.Row(i) + head * 3 * qkv_dim + 2 * qkv_dim;
MulByConstAndAdd(head_att[i], v, att_out, qkv_dim);
}
});
}
// Sums encoded (`att_out`) over num_heads (`layer_config_.heads`) and
// head_dim (`qkv_dim`) into output (`att_sums`).
HWY_NOINLINE void SumHeads() {
PROFILER_ZONE("Gen.VitAttention.SumHeads");
auto* bias = layer_weights_.vit.attn_out_b.PackedScale1();
// att_weights and att_out are concatenated heads, each of length
// qkv_dim. Thus the [num_tokens_, layer_config_.model_dim]
// matmul output is the sum over heads.
auto att_sums = RowPtrFromMat(allocator_, activations_.att_sums);
MatMul(activations_.att_out, layer_weights_.vit.attn_out_w, bias,
*activations_.env, att_sums);
}
public:
VitAttention(size_t num_tokens, size_t layer, Activations& activations,
const LayerWeightsPtrs<T>* layer_weights)
: num_tokens_(num_tokens),
layer_(layer),
activations_(activations),
layer_weights_(*layer_weights),
layer_config_(layer_weights->layer_config),
allocator_(activations.env->ctx.allocator),
pool_(activations.env->ctx.pools.Pool(0)) {}
HWY_INLINE void operator()() {
ComputeQKV();
if (activations_.weights_config.wrapping == PromptWrapping::GEMMA_VLM) {
DotSoftmaxWeightedSumMatrix();
} else {
DotSoftmaxWeightedSum();
}
SumHeads();
}
private:
const size_t num_tokens_;
const size_t layer_;
Activations& activations_;
const LayerWeightsPtrs<T>& layer_weights_;
const LayerConfig& layer_config_;
const Allocator& allocator_;
hwy::ThreadPool& pool_;
};
template <typename T>
HWY_NOINLINE void Activation(ActivationType activation, T* HWY_RESTRICT c1,
const T* HWY_RESTRICT c2, size_t count) {
PROFILER_ZONE("Gen.Activation");
namespace hn = hwy::HWY_NAMESPACE;
using DF = hn::ScalableTag<T>;
using VF = hn::Vec<DF>;
// ActivationType::Gelu
if (c2 == nullptr) { // No multiplier, just Gelu.
Gelu(c1, count);
return;
};
// Has multiplier, Gelu(c1) * c2.
hn::Transform1(DF(), c1, count, c2, [](DF df, VF v, VF mul) HWY_ATTR {
return hn::Mul(mul, Gelu(df, v));
});
}
// No C2 multiplier.
template <class Mat>
void ActivationBatched(ActivationType activation, Mat& c1) {
using T = typename Mat::T;
for (size_t i = 0; i < c1.Rows(); ++i) {
// Cast to correct type so type deduction works.
Activation(activation, c1.Row(i), static_cast<const T*>(nullptr),
c1.Cols());
}
}
template <class Mat>
void ActivationBatched(ActivationType activation, Mat& c1, const Mat* c2) {
using T = typename Mat::T;
HWY_DASSERT(c1.SameShape(*c2));
if (c2 && c2->HasPtr()) {
for (size_t i = 0; i < c1.Rows(); ++i) {
Activation(activation, c1.Row(i), c2->Row(i), c1.Cols());
}
} else { // No multiplier
for (size_t i = 0; i < c1.Rows(); ++i) {
Activation(activation, c1.Row(i), static_cast<const T*>(nullptr),
c1.Cols());
}
}
}
template <typename T>
HWY_NOINLINE void FFWNoVit(Activations& activations,
const LayerWeightsPtrs<T>* layer_weights) {
PROFILER_ZONE("Gen.FFW");
const size_t ffh_hidden_dim = layer_weights->layer_config.ff_hidden_dim;
const bool add_bias = layer_weights->layer_config.ff_biases;
const float* bias1 =
add_bias ? layer_weights->ffw_gating_biases.PackedScale1() : nullptr;
const float* bias2 = add_bias ? bias1 + ffh_hidden_dim : nullptr;
const float* output_bias =
add_bias ? layer_weights->ffw_output_biases.PackedScale1() : nullptr;
// Define slightly more readable names for the weights and activations.
const Allocator& allocator = activations.env->ctx.allocator;
auto hidden_activations = RowPtrFromMat(allocator, activations.C1);
auto multiplier = RowPtrFromMat(allocator, activations.C2);
auto ffw_out = RowPtrFromMat(allocator, activations.ffw_out);
using WeightT = typename decltype(layer_weights->gating_einsum_w)::T;
// gating_einsum_w holds two half-matrices. We plan to change the importer to
// avoid this confusion by splitting into gating_einsum_w1 and
// gating_einsum_w2. TODO: move into Reshape().
const bool split = layer_weights->gating_einsum_w.HasPtr();
ConstMat<WeightT> w1(split ? layer_weights->gating_einsum_w
: layer_weights->gating_einsum_w1);
ConstMat<WeightT> w2(split ? layer_weights->gating_einsum_w
: layer_weights->gating_einsum_w2);
if (split) {
w2.ofs = w2.Row(ffh_hidden_dim);
// Ensure that B.Extents().row matches C.Cols() because MatMul checks that.
w1.ShrinkRows(ffh_hidden_dim);
w2.ShrinkRows(ffh_hidden_dim);
}
// Compute the hidden layer activations.
MatMul(activations.pre_ffw_rms_out, w1, bias1, *activations.env,
hidden_activations);
MatMul(activations.pre_ffw_rms_out, w2, bias2, *activations.env, multiplier);
// Activation (Gelu) and maybe multiply by gate. Store activations in act.
ActivationBatched(layer_weights->layer_config.activation, activations.C1,
&activations.C2);
// Hidden layer -> output layer.
MatMul(activations.C1, layer_weights->linear_w, output_bias, *activations.env,
ffw_out);
}
// Same as FFWNoVit, but with different layer_weights members and no second
// gating matrix.
template <typename T>
HWY_NOINLINE void FFWVit(Activations& activations,
const LayerWeightsPtrs<T>* layer_weights) {
PROFILER_ZONE("Gen.FFW.ViT");
const bool add_bias = layer_weights->layer_config.ff_biases;
const float* bias1 =
add_bias ? layer_weights->vit.linear_0_b.PackedScale1() : nullptr;
const float* output_bias =
add_bias ? layer_weights->vit.linear_1_b.PackedScale1() : nullptr;
// Define slightly more readable names for the weights and activations.
const Allocator& allocator = activations.env->ctx.allocator;
auto hidden_activations = RowPtrFromMat(allocator, activations.C1);
auto ffw_out = RowPtrFromMat(allocator, activations.ffw_out);
// Compute the hidden layer activations.
MatMul(activations.pre_ffw_rms_out, layer_weights->vit.linear_0_w, bias1,
*activations.env, hidden_activations);
// Activation (Gelu), store in act.
RowPtrF multiplier = RowPtrF(allocator, nullptr, 0);
ActivationBatched(layer_weights->layer_config.activation, activations.C1);
// Hidden layer -> output layer.
MatMul(activations.C1, layer_weights->vit.linear_1_w, output_bias,
*activations.env, ffw_out);
}
// `batch_idx` indicates which row of `x` to write to.
// `pos` is the *token*'s position, not the start of the batch, because this is
// called for batches of tokens in prefill, but batches of queries in decode.
//
// For GEMMA_VLM, image tokens are copied into -2 locations (per the Gemma 3
// spec) until we run out of image tokens. This allows for a multi-image prompt
// if -2 locations with appropriate begin/end image tokens are created by the
// calling application.
template <typename T>
HWY_NOINLINE void EmbedMMToken(int token, size_t batch_idx, size_t pos,
size_t pos_in_prompt,
const ModelWeightsPtrs<T>& weights,
MatStorageT<float>& x,
const ImageTokens* image_tokens,
size_t& image_token_position) {
// Image tokens just need to be copied.
if (weights.weights_config.wrapping == PromptWrapping::GEMMA_VLM &&
image_tokens != nullptr && token == -2 &&
image_token_position < image_tokens->Rows()) {
hwy::CopyBytes(image_tokens->Row(image_token_position), x.Row(batch_idx),
x.Cols() * x.ElementBytes());
image_token_position++;
return;
}
if (weights.weights_config.wrapping == PromptWrapping::PALIGEMMA &&
image_tokens != nullptr && pos_in_prompt < image_tokens->Rows()) {
hwy::CopyBytes(image_tokens->Row(pos_in_prompt), x.Row(batch_idx),
x.Cols() * x.ElementBytes());
return;
}
const size_t model_dim = weights.weights_config.model_dim;
const size_t vocab_size = weights.weights_config.vocab_size;
const float emb_scaling = EmbeddingScaling(model_dim);
HWY_DASSERT(token >= 0);
HWY_DASSERT(token < static_cast<int>(vocab_size));
const hn::ScalableTag<float> df;
// Using `Stride` to compute the offset works for both NUQ (because we use an
// offset and NUQ is never padded) and padded, because non-NUQ types are
// seekable, hence the offset can also skip any padding.
const size_t embedding_ofs =
token * weights.embedder_input_embedding.Stride();
HWY_ASSERT(weights.embedder_input_embedding.Cols() == model_dim);
const auto embedding_span = MakeSpan(weights.embedder_input_embedding.Row(0),
embedding_ofs + model_dim);
DecompressAndZeroPad(df, embedding_span, embedding_ofs, x.Row(batch_idx),
model_dim);
MulByConst(emb_scaling * weights.embedder_input_embedding.Scale(),
x.Row(batch_idx), model_dim);
if (weights.weights_config.absolute_pe) {
AddAbsolutePositionalEmbeddings(x.Row(batch_idx), model_dim, pos);
}
}
// `batch_idx` indicates which row of `x` to write to.
// `pos` is the *token*'s position, not the start of the batch, because this is
// called for batches of tokens in prefill, but batches of queries in decode.
// This version of the function doesn't track internal image token position.
template <typename T>
HWY_NOINLINE void EmbedToken(int token, size_t batch_idx, size_t pos,
size_t pos_in_prompt,
const ModelWeightsPtrs<T>& weights,
MatStorageT<float>& x,
const ImageTokens* image_tokens) {
size_t image_token_position = 0;
EmbedMMToken<T>(token, batch_idx, pos, pos_in_prompt, weights, x,
image_tokens, image_token_position);
}
template <typename T2, class LayerWeights>
HWY_NOINLINE void ResidualConnection(const MatPtrT<T2>& other,
MatPtrT<float>& HWY_RESTRICT x,
const LayerWeights* layer_weights,
bool is_attention) {
// ResidualType::Add
AddFromBatched(other, x);
}
template <typename WeightT, typename InOutT>
void PostNorm(PostNormType post_norm, const MatPtrT<WeightT>& weights,
MatPtrT<InOutT>& inout) {
HWY_DASSERT(weights.Rows() == 1);
if (post_norm == PostNormType::Scale) {
RMSNormInplaceBatched(weights, inout);
}
}
template <typename T>
HWY_NOINLINE void TransformerLayer(const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end,
size_t num_tokens, size_t cache_layer_idx,
const LayerWeightsPtrs<T>* layer_weights,
Activations& activations,
const hwy::Divisor& div_seq_len,
const KVCaches& kv_caches) {
auto type = layer_weights->layer_config.type;
RMSNormBatched(activations.x, layer_weights->pre_attention_norm_scale,
activations.pre_att_rms_out);
Attention(type, queries_pos, queries_prefix_end, num_tokens, cache_layer_idx,
activations, layer_weights, div_seq_len, kv_caches);
PostNorm(layer_weights->layer_config.post_norm,
layer_weights->post_attention_norm_scale, activations.att_sums);
ResidualConnection(activations.att_sums, activations.x, layer_weights,
/*is_attention=*/true);
RMSNormBatched(activations.x, layer_weights->pre_ffw_norm_scale,
activations.pre_ffw_rms_out);
if (layer_weights->layer_config.type == LayerAttentionType::kVit) {
FFWVit(activations, layer_weights);
} else {
FFWNoVit(activations, layer_weights);
}
PostNorm(layer_weights->layer_config.post_norm,
layer_weights->post_ffw_norm_scale, activations.ffw_out);
ResidualConnection(activations.ffw_out, activations.x, layer_weights,
/*is_attention=*/false);
}
// Vit transformer layer. Some comments below refer to the Vit implementation in
// the Big Vision codebase. See
// github.com/google-research/big_vision/blob/main/big_vision/models/vit.py
// TODO(keysers): consider adding a wrapper for both LayerNorm with RMSNorm and
// try merging this with TransformerLayer.
template <typename T>
HWY_NOINLINE void VitTransformerLayer(size_t num_tokens, size_t layer,
const LayerWeightsPtrs<T>* layer_weights,
Activations& activations) {
const size_t model_dim = activations.weights_config.model_dim;
auto type = layer_weights->layer_config.type;
HWY_DASSERT(type == LayerAttentionType::kVit);
(void)type;
(void)model_dim;
auto& x = activations.x;
HWY_DASSERT(x.Rows() == num_tokens);
HWY_DASSERT(x.Cols() == model_dim);
// y = nn.LayerNorm()(x)
// y ~ pre_att_rms_out
LayerNormBatched(x, layer_weights->vit.layer_norm_0_scale,
layer_weights->vit.layer_norm_0_bias,
activations.pre_att_rms_out);
// y = out["sa"] = nn.MultiHeadDotProductAttention(...)(y, y)
// y ~ att_sums
VitAttention<T>(num_tokens, layer, activations, layer_weights)();
// x = out["+sa"] = x + y
AddFromBatched(activations.att_sums, x);
// y = nn.LayerNorm()(x)
// y ~ pre_ffw_rms_out
LayerNormBatched(x, layer_weights->vit.layer_norm_1_scale,
layer_weights->vit.layer_norm_1_bias,
activations.pre_ffw_rms_out);
// y = out["mlp"] = MlpBlock(...)(y)
// y ~ ffw_out
FFWVit(activations, layer_weights);
// x = out["+mlp"] = x + y
AddFromBatched(activations.ffw_out, x);
}
// Prefill() and Transformer() increment positions in-place.
using QueriesMutablePos = hwy::Span<size_t>;
// Populates KV cache for batches of tokens from one query at a time.
template <typename T>
HWY_NOINLINE void Prefill(
const QueriesPromptTokens& queries_prompt,
const QueriesMutablePos& queries_pos, const QueriesPos& queries_prefix_end,
const size_t query_idx_start, const ModelWeightsPtrs<T>& weights,
Activations& activations, const RuntimeConfig& runtime_config,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches) {
PROFILER_ZONE("Gen.Prefill");
const size_t num_queries = queries_prompt.size();
HWY_DASSERT(queries_pos.size() == num_queries);
HWY_DASSERT(queries_prefix_end.size() == num_queries);
HWY_DASSERT(kv_caches.size() == num_queries);
// Batches are important for amortizing loading weights over multiple tokens.
// This is possible in prefill because we know all tokens beforehand, whereas
// decode depends on the previous output token. However, each prefill batch of
// a query requires that preceding batches already wrote to the KV cache,
// hence we sequentially loop over token batches. We can reduce the number of
// iterations by increasing the batch size, but this also increases arithmetic
// intensity, and so we are eventually compute-limited. We could devote some
// threads to parallelizing over queries, but for simplicity we assign them
// all to MatMul.
const size_t max_tbatch_size = activations.x.Rows();
// For each query. `qi` is within the batch, not the global query index.
for (size_t qi = 0; qi < num_queries; ++qi) {
// Single query at a time, so pass slices of the spans because
// GemmaAttention will only access the first KV cache and position.
QueriesPos single_query_pos(&queries_pos[qi], 1);
QueriesPos single_query_prefix_end(&queries_prefix_end[qi], 1);
KVCaches single_kv_cache(&kv_caches[qi], 1);
const size_t prompt_size = queries_prompt[qi].size();
// In autoregressive mode, we don't need to prefill the last token, so - 1.
size_t prefill_this_query = prompt_size - 1;
const size_t prefix_end_this_query = queries_prefix_end[qi];
// We can't attend beyond the prompt_size.
HWY_ASSERT(prefix_end_this_query <= prompt_size);
// Special case: if the prefix includes the last token, we need to prefill
// the last token, too. However, we need to rewind this for the generation
// of the first token. So we need to keep track of this.
// TODO: consider implementing masking instead of this logic?
const bool attend_to_last_token =
(prefill_this_query < prefix_end_this_query);
if (attend_to_last_token) {
// The difference can be at most 1.
prefill_this_query += 1;
HWY_ASSERT(prefill_this_query == prefix_end_this_query);
}
// In prefix-LM mode, we need to look at all the tokens for the prefix in
// one iteration through the layers, so we need a large enough batch size.
HWY_ASSERT(prefix_end_this_query == 0 ||
max_tbatch_size >= prefill_this_query);
// For each batch of tokens in the query:
for (size_t tbatch_start = 0; tbatch_start < prefill_this_query;
tbatch_start += max_tbatch_size) {
const size_t tbatch_size =
HWY_MIN(max_tbatch_size, prefill_this_query - tbatch_start);
activations.SetBatchSize(tbatch_size);
// Fill activations.x (much faster than TransformerLayer).
size_t image_token_position = 0;
for (size_t ti = 0; ti < tbatch_size; ++ti) {
const size_t pos = queries_pos[qi] + ti;
const size_t pos_in_prompt = tbatch_start + ti;
const int token = queries_prompt[qi][pos_in_prompt];
EmbedMMToken(token, ti, pos, pos_in_prompt, weights, activations.x,
runtime_config.image_tokens, image_token_position);
}
// Transformer with one batch of tokens from a single query.
for (size_t layer = 0;
layer < weights.weights_config.layer_configs.size(); ++layer) {
const LayerWeightsPtrs<T>* layer_weights = weights.GetLayer(layer);
TransformerLayer(single_query_pos, single_query_prefix_end, tbatch_size,
layer, layer_weights, activations, div_seq_len,
single_kv_cache);
}
// NOTE: we unconditionally call StreamToken, even if EOS.
for (size_t ti = 0; ti < tbatch_size; ++ti) {
const size_t pos = queries_pos[qi] + ti;
const size_t pos_in_prompt = tbatch_start + ti;
const int token = queries_prompt[qi][pos_in_prompt];
if (pos_in_prompt < prompt_size - 1) {
runtime_config.StreamToken(query_idx_start + qi, pos, token, 0.0f);
} else {
// The last token will be streamed later and we should only get here
// if we need to attend to the last token because it is in the prefix.
HWY_ASSERT(attend_to_last_token);
}
}
queries_pos[qi] += tbatch_size;
} // for tbatch_start
if (attend_to_last_token) {
// We need to rewind the position for the last token that we only
// attended to to make sure the prefix LM sees everything.
// This means we duplicate work on the last prompt token in autoregressive
// decoding. Alternatives: (1) real masking; (2) always prefill the last
// token and only generate the next one from the already prefilled
// activations.
queries_pos[qi] -= 1;
}
}
}
// Gets the patches of the image and embeds them with the image embedding
// kernel. The result is stored in activations.x.
template <typename T>
HWY_NOINLINE void EmbedImagePatches(const Image& image,
const ModelWeightsPtrs<T>& weights,
Activations& activations) {
const size_t model_dim = weights.weights_config.vit_config.model_dim;
const size_t patch_width = weights.weights_config.vit_config.patch_width;
const size_t seq_len = weights.weights_config.vit_config.seq_len;
const size_t patch_size = patch_width * patch_width * 3;
HWY_DASSERT(weights.vit_img_embedding_kernel.Rows() == model_dim);
HWY_DASSERT(weights.vit_img_embedding_kernel.Cols() == patch_size);
HWY_DASSERT(activations.x.Cols() == model_dim);
std::vector<hwy::AlignedFreeUniquePtr<float[]>> image_patches(seq_len);
for (size_t i = 0; i < seq_len; ++i) {
image_patches[i] = hwy::AllocateAligned<float>(patch_size);
image.GetPatch(i, image_patches[i].get());
}
// img/embedding/kernel has original shape (14, 14, 3, 1152)
// H x W x C x D transposed to D x (H x W x C) so here (1152, 14 * 14 * 3)
// image_patches is (256, 14 * 14 * 3)
// This could be done as one MatMul like:
// MatStorageT<float> image_patches("patches", Extents2D(kSeqLen,
// kPatchSize), MatPadding::kPacked);
// [Get patches]
// MatMul(
// MatFromBatch(kVitSeqLen, image_patches),
// MatFromWeights(weights.vit_img_embedding_kernel),
// weights.vit_img_embedding_bias.PackedScale1(), *activations.env,
// RowPtrF(activations.x.Row(0), kVitModelDim));
// However, MatMul currently requires that
// A.cols % (2 * hn::Lanes(hn::ScalableTag<MulT>())) == 0
// which is not the case here. We should relax that requirement on MatMul and
// then use the above. For now, we rely on MatVecAdd instead.
for (size_t i = 0; i < seq_len; ++i) {
MatVecAdd(weights.vit_img_embedding_kernel, 0, model_dim, patch_size,
image_patches[i].get(),
weights.vit_img_embedding_bias.PackedScale1(),
activations.x.Row(i), activations.env->ctx.pools.Pool(0));
}
// Add position embeddings.
AddFromBatched(weights.vit_img_pos_embedding, activations.x);
}
// Prefills the image tokens with the ViT encoder.
template <typename T>
HWY_NOINLINE void PrefillVit(const ModelWeightsPtrs<T>& weights,
const RuntimeConfig& runtime_config,
const Image& image, ImageTokens& image_tokens,
Activations& activations) {
PROFILER_ZONE("Gen.PrefillVit");
const size_t num_tokens = weights.weights_config.vit_config.seq_len;
const size_t vit_model_dim = weights.weights_config.vit_config.model_dim;
HWY_ASSERT(num_tokens == activations.x.Rows());
// Embed the image patches.
EmbedImagePatches(image, weights, activations);
// Go through all layers.
for (size_t layer = 0;
layer < weights.weights_config.vit_config.layer_configs.size();
++layer) {
const LayerWeightsPtrs<T>* layer_weights = weights.VitLayer(layer);
VitTransformerLayer(num_tokens, layer, layer_weights, activations);
}
// Final Layernorm.
LayerNormBatched(activations.x, weights.vit_encoder_norm_scale,
weights.vit_encoder_norm_bias, activations.x);
if (weights.weights_config.wrapping == PromptWrapping::GEMMA_VLM) {
activations.x = AvgPool4x4(activations.x);
// Apply soft embedding norm before input projection.
RMSNormInplace(weights.mm_embed_norm.PackedScale1(), 0,
activations.x.Row(0), vit_model_dim);
}
// Apply head embedding into image_tokens of size of the LLM kModelDim.
MatMul(activations.x, weights.vit_img_head_kernel,
weights.vit_img_head_bias.PackedScale1(), *activations.env,
RowPtrFromMat(activations.env->ctx.allocator, image_tokens));
}
// Generates one token for each query. `queries_token` is the previous token
// from each query, and `queries_pos` are their position in the sequence.
template <typename T>
HWY_NOINLINE void Transformer(
const QueriesToken& queries_token, const QueriesMutablePos& queries_pos,
const QueriesPos& queries_prefix_end, const ModelWeightsPtrs<T>& weights,
Activations& activations, const hwy::Divisor& div_seq_len,
const KVCaches& kv_caches, const LayersOutputFunc& layers_output,
const ActivationsObserverFunc& activations_observer) {
const size_t num_queries = queries_token.size();
HWY_DASSERT(queries_pos.size() == num_queries);
HWY_DASSERT(queries_prefix_end.size() == num_queries);
if (layers_output) {
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
const float token_f = queries_token[query_idx];
layers_output(query_idx, queries_pos[query_idx], "tokens", -1, &token_f,
1);
}
}
size_t image_token_position = 0;
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
EmbedMMToken(queries_token[query_idx], query_idx, queries_pos[query_idx],
/*pos_in_prompt=*/0, weights, activations.x,
/*image_tokens=*/nullptr, image_token_position);
}
for (size_t layer = 0; layer < weights.c_layers.size(); ++layer) {
const LayerWeightsPtrs<T>* layer_weights = weights.GetLayer(layer);
TransformerLayer(queries_pos, queries_prefix_end, /*num_tokens=*/1, layer,
layer_weights, activations, div_seq_len, kv_caches);
if (activations_observer) {
activations_observer(queries_pos, layer, activations);
}
}
RMSNormInplaceBatched(weights.final_norm_scale, activations.x);
if (activations_observer) {
activations_observer(queries_pos, -1, activations);
}
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
queries_pos[query_idx] += 1;
}
}
// Placeholder for internal test3, do not remove
// Returns the min and max number of tokens for all queries.
static size_t MaxQueryLength(const QueriesPromptTokens& queries_prompt) {
size_t max_prompt_size = 0;
for (size_t i = 0; i < queries_prompt.size(); ++i) {
max_prompt_size = std::max(max_prompt_size, queries_prompt[i].size());
}
return max_prompt_size;
}
// Holds "is at end of stream" state for each query.
class TokenStreamer {
public:
explicit TokenStreamer(const RuntimeConfig& runtime_config,
const ModelConfig& model_config)
: runtime_config_(runtime_config), model_config_(model_config) {}
// Returns whether the query was already at, or has just reached, the end of
// the stream: either via token == eos_id, or StreamToken returning false.
bool operator()(size_t query_idx, size_t pos, int token, float prob) {
if (HWY_UNLIKELY(is_eos_.Get(query_idx))) return true;
if (!runtime_config_.StreamToken(query_idx, pos, token, prob) ||
model_config_.IsEOS(token)) {
is_eos_.Set(query_idx);
return true;
}
return false;
}
private:
const RuntimeConfig& runtime_config_;
const ModelConfig& model_config_;
hwy::BitSet4096<> is_eos_;
};
HWY_INLINE SampleFunc ChooseSampleFunc(const RuntimeConfig& runtime_config) {
// If user provided a sample_func, use it.
if (runtime_config.sample_func) return runtime_config.sample_func;
// Fast path for top-1 with no accept_token.
if (runtime_config.top_k == 1 && !runtime_config.accept_token) {
return [](float* logits, size_t vocab_size) HWY_ATTR -> TokenAndProb {
PROFILER_ZONE("Gen.Sample Top1");
return Top1OfSoftmax(logits, vocab_size);
};
}
// General case: Softmax with top-k sampling.
return [&runtime_config](float* logits,
size_t vocab_size) HWY_ATTR -> TokenAndProb {
PROFILER_ZONE("Gen.Sample general");
return FusedSoftmaxAndSampleTopK(
logits, runtime_config.top_k, vocab_size, *runtime_config.gen,
runtime_config.temperature, runtime_config.accept_token);
};
}
template <typename T>
// Runs one decode step for all the queries in the batch. Returns true if all
// queries are at <end_of_sentence>.
bool DecodeStepT(const ModelConfig& config, const ModelWeightsPtrs<T>& weights,
const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt,
const size_t query_idx_start, const KVCaches& kv_caches,
const QueriesPos& queries_prefix_end,
const hwy::Divisor div_seq_len, const size_t vocab_size,
const SampleFunc& sample_token, Activations& activations,
TokenStreamer& token_streamer, std::vector<int>& gen_tokens,
TimingInfo& timing_info,
const QueriesMutablePos& queries_mutable_pos) {
const size_t num_queries = queries_prompt.size();
// Decode generates one token per query and increments
// queries_mutable_pos.
Transformer(QueriesToken(gen_tokens.data(), num_queries), queries_mutable_pos,
queries_prefix_end, weights, activations, div_seq_len, kv_caches,
runtime_config.layers_output,
runtime_config.activations_observer);
// queries_pos are incremented by Transformer.
HWY_DASSERT(num_queries == activations.x.Rows());
bool all_queries_eos = true;
{
PROFILER_ZONE("Gen.EmbeddingMatmul");
// Compute logits from last layer activations.
MatMul(activations.x, weights.embedder_input_embedding,
/*add=*/nullptr, *activations.env,
RowPtrFromMat(activations.env->ctx.allocator, activations.logits));
}
PROFILER_ZONE("Gen.Softcap+Sample+Stream");
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
float* HWY_RESTRICT logits = activations.logits.Row(query_idx);
MaybeLogitsSoftCap(weights.weights_config.final_cap, logits, vocab_size);
const TokenAndProb tp = sample_token(logits, vocab_size);
timing_info.NotifyGenerated();
const bool is_eos =
token_streamer(query_idx_start + query_idx,
queries_mutable_pos[query_idx], tp.token, tp.prob);
all_queries_eos &= is_eos;
gen_tokens[query_idx] = is_eos ? config.eos_id : tp.token;
}
return all_queries_eos;
}
// Generates one continuation for each query in `queries_prompt`, which is one
// qbatch whose size is at most the `batch_size` passed to
// `activations.Allocate`.
//
// `queries_pos` stores the KV cache position for each query. In the first turn
// of a chat, pos = 0; we increment each query's position after each token.
//
// `query_idx_start` is the query_idx of the first query in the batch, so that
// `StreamFunc` gets the global query index, not relative to the batch.
//
// `kv_caches` is for the batch, size must match `queries_prompt`.
template <typename T>
void GenerateT(const ModelStore& model, const ModelWeightsPtrs<T>& weights,
Activations& activations, const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt,
const QueriesPos& queries_pos_in,
const QueriesPos& queries_prefix_end,
const size_t query_idx_start, const KVCaches& kv_caches,
TimingInfo& timing_info) {
// Griffin assumes that the recurrent block cache is zero-initialized.
for (size_t i = 0; i < kv_caches.size(); ++i) {
if (queries_pos_in[i] == 0) {
kv_caches[i].ZeroGriffinCache(); // No-op for non-Griffin models.
}
}
// Copy so we can increment without requiring users to pass in a mutable span.
std::vector<size_t> queries_pos_copy(queries_pos_in.cbegin(),
queries_pos_in.cend());
const QueriesMutablePos queries_mutable_pos(queries_pos_copy.data(),
queries_pos_copy.size());
// Sanity check: prompts should not be empty, nor start with EOS.
for (size_t query_idx = 0; query_idx < queries_prompt.size(); ++query_idx) {
const PromptTokens& prompt = queries_prompt[query_idx];
HWY_ASSERT(prompt.size() != 0 && prompt[0] != model.Config().eos_id);
}
const size_t num_queries = queries_prompt.size();
HWY_ASSERT(num_queries <= 4096); // TokenStreamer uses BitSet4096.
HWY_ASSERT(num_queries <= activations.x.Rows());
HWY_ASSERT(queries_pos_in.size() == num_queries);
HWY_ASSERT(kv_caches.size() == num_queries);
const hwy::Divisor div_seq_len(static_cast<uint32_t>(kv_caches[0].seq_len));
size_t max_prompt_size = MaxQueryLength(queries_prompt);
size_t max_generated_tokens = runtime_config.max_generated_tokens;
RangeChecks(weights.weights_config, max_generated_tokens, max_prompt_size);
const SampleFunc sample_token = ChooseSampleFunc(runtime_config);
// Prefill stops before min_prompt_size - 1 because the last prompt
// token is the first input token for generation.
timing_info.prefill_start = hwy::platform::Now();
// If tbatch is larger than the qbatch we already have in `activations`, then
// allocate prefill_activations, otherwise reuse.
const bool use_prefill_activations =
runtime_config.prefill_tbatch_size > activations.x.Rows();
Activations prefill_activations(
weights.weights_config,
use_prefill_activations ? runtime_config.prefill_tbatch_size : 0,
activations.env);
Prefill(queries_prompt, queries_mutable_pos, queries_prefix_end,
query_idx_start, weights,
use_prefill_activations ? prefill_activations : activations,
runtime_config, div_seq_len, kv_caches);
// Compute the number of tokens that were prefilled and notify timing_info.
size_t prefilled_tokens = 0;
for (size_t qi = 0; qi < num_queries; ++qi) {
prefilled_tokens += queries_prompt[qi].size() - 1;
}
timing_info.NotifyPrefill(prefilled_tokens);
// queries_pos are incremented by Prefill.
// Storage for the last generated token from each query, passed to the next
// Transformer() call.
std::vector<int> gen_tokens(num_queries);
// Stream the last prompt token from each query and fill gen_tokens.
TokenStreamer token_streamer(runtime_config, model.Config());
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
size_t last_token_pos_in_prompt =
queries_mutable_pos[query_idx] - queries_pos_in[query_idx];
gen_tokens[query_idx] = queries_prompt[query_idx][last_token_pos_in_prompt];
(void)token_streamer(query_idx_start + query_idx,
queries_mutable_pos[query_idx], gen_tokens[query_idx],
0.0f);
}
{
const size_t vocab_size = model.Config().vocab_size;
timing_info.generate_start = hwy::platform::Now();
for (size_t gen = 0; gen < max_generated_tokens; ++gen) {
bool all_queries_eos = DecodeStepT<T>(
model.Config(), weights, runtime_config, queries_prompt,
query_idx_start, kv_caches, queries_prefix_end, div_seq_len,
vocab_size, sample_token, activations, token_streamer, gen_tokens,
timing_info, queries_mutable_pos);
if (all_queries_eos) break;
} // foreach token to generate
timing_info.NotifyGenerateDone();
}
}
template <typename T>
void GenerateSingleT(const ModelStore& model,
const ModelWeightsPtrs<T>& weights,
const RuntimeConfig& runtime_config,
const PromptTokens& prompt, size_t pos, size_t prefix_end,
KVCache& kv_cache, MatMulEnv* env,
TimingInfo& timing_info) {
constexpr size_t kNumQueries = 1;
const size_t qbatch_start = 0;
// TODO: move into Gemma?
Activations activations(model.Config(), kNumQueries, env);
const QueriesPromptTokens queries_prompt(&prompt, kNumQueries);
QueriesPos queries_pos(&pos, kNumQueries);
const QueriesPos queries_prefix_end(&prefix_end, kNumQueries);
const KVCaches kv_caches{&kv_cache, kNumQueries};
GenerateT<T>(model, weights, activations, runtime_config, queries_prompt,
queries_pos, queries_prefix_end, qbatch_start, kv_caches,
timing_info);
}
template <typename T>
void GenerateBatchT(const ModelStore& model,
const ModelWeightsPtrs<T>& weights,
const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt,
const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end,
const KVCaches& kv_caches, MatMulEnv* env,
TimingInfo& timing_info) {
const size_t num_queries = queries_prompt.size();
HWY_ASSERT(queries_pos.size() == num_queries);
HWY_ASSERT(kv_caches.size() >= num_queries);
// Griffin does not support query batching.
size_t max_qbatch_size = runtime_config.decode_qbatch_size;
for (const LayerConfig& layer_config : model.Config().layer_configs) {
if (layer_config.type == LayerAttentionType::kGriffinRecurrentBlock) {
max_qbatch_size = 1;
break;
}
}
Activations activations(model.Config(), max_qbatch_size, env);
for (size_t qbatch_start = 0; qbatch_start < num_queries;
qbatch_start += max_qbatch_size) {
// Generate one batch of tokens from `qbatch_size` queries.
const size_t qbatch_size =
HWY_MIN(num_queries - qbatch_start, max_qbatch_size);
activations.SetBatchSize(qbatch_size);
const QueriesPromptTokens qbatch_prompts(&queries_prompt[qbatch_start],
qbatch_size);
QueriesPos qbatch_pos(&queries_pos[qbatch_start], qbatch_size);
const QueriesPos qbatch_prefix_end(&queries_prefix_end[qbatch_start],
qbatch_size);
const KVCaches qbatch_kv(&kv_caches[qbatch_start], qbatch_size);
GenerateT<T>(model, weights, activations, runtime_config, qbatch_prompts,
qbatch_pos, qbatch_prefix_end, qbatch_start, qbatch_kv,
timing_info);
}
}
template <typename T>
void GenerateImageTokensT(const ModelStore& model,
const ModelWeightsPtrs<T>& weights,
const RuntimeConfig& runtime_config,
const Image& image, ImageTokens& image_tokens,
MatMulEnv* env) {
if (model.Config().vit_config.layer_configs.empty()) {
HWY_ABORT("Model does not support generating image tokens.");
}
RuntimeConfig prefill_runtime_config = runtime_config;
ModelConfig vit_config = GetVitConfig(model.Config());
prefill_runtime_config.prefill_tbatch_size =
vit_config.seq_len / (vit_config.pool_dim * vit_config.pool_dim);
Activations prefill_activations(vit_config, vit_config.seq_len, env);
// Weights are for the full PaliGemma model, not just the ViT part.
PrefillVit(weights, prefill_runtime_config, image, image_tokens,
prefill_activations);
}
} // namespace HWY_NAMESPACE
#if HWY_ONCE
// These are extern functions defined by instantiations/*.cc, which include this
// 'header' after defining `GEMMA_TYPE`.
void GenerateSingle( // NOLINT(misc-definitions-in-headers)
const ModelStore& model, const ModelWeightsPtrs<GEMMA_TYPE>& weights,
const RuntimeConfig& runtime_config, const PromptTokens& prompt, size_t pos,
size_t prefix_end, KVCache& kv_cache, MatMulEnv* env,
TimingInfo& timing_info) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateSingleT<GEMMA_TYPE>)
(model, weights, runtime_config, prompt, pos, prefix_end, kv_cache, env,
timing_info);
}
void GenerateBatch( // NOLINT(misc-definitions-in-headers)
const ModelStore& model, const ModelWeightsPtrs<GEMMA_TYPE>& weights,
const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt, const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end, const KVCaches& kv_caches,
MatMulEnv* env, TimingInfo& timing_info) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateBatchT<GEMMA_TYPE>)
(model, weights, runtime_config, queries_prompt, queries_pos,
queries_prefix_end, kv_caches, env, timing_info);
}
void GenerateImageTokens( // NOLINT(misc-definitions-in-headers)
const ModelStore& model, const ModelWeightsPtrs<GEMMA_TYPE>& weights,
const RuntimeConfig& runtime_config, const Image& image,
ImageTokens& image_tokens, MatMulEnv* env) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateImageTokensT<GEMMA_TYPE>)
(model, weights, runtime_config, image, image_tokens, env);
}
#endif // HWY_ONCE
} // namespace gcpp
HWY_AFTER_NAMESPACE();
#endif // THIRD_PARTY_GEMMA_CPP_GEMMA_GEMMA_INL_H_
Reference in New Issue View Git Blame Copy Permalink