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

1421 lines
61 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 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 <stddef.h>
#include <stdio.h>
#include <algorithm> // std::min
#include <string>
#include <type_traits>
#include <vector>
#include "gemma/activations.h"
#include "gemma/common.h"
#include "gemma/configs.h"
#include "gemma/gemma.h"
#include "gemma/weights.h"
// Placeholder for internal test4, do not remove
#include "ops/matmul-inl.h"
#include "ops/matvec-inl.h"
#include "ops/ops-inl.h"
#include "paligemma/image.h"
#include "util/allocator.h"
#include "util/threading.h"
#include "hwy/aligned_allocator.h"
#include "hwy/base.h"
#include "hwy/bit_set.h"
#include "hwy/contrib/thread_pool/thread_pool.h"
#include "hwy/highway.h"
#include "hwy/profiler.h"
#include "hwy/timer.h"
#ifndef GEMMA_CONFIG
#if HWY_IDE
// Provide a definition so the IDE does not complain.
#define GEMMA_CONFIG ConfigGemmaTiny<float>
#else
#error "Only include from instantiations/*.cc, which must define GEMMA_CONFIG"
#endif // HWY_IDE
#endif // GEMMA_CONFIG
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 <class TConfig>
HWY_NOINLINE void GriffinRecurrent(
size_t batch_start, size_t num_tokens, size_t layer,
Activations& activations, const CompressedLayer<TConfig>* layer_weights,
const KVCaches& kv_caches) {
PROFILER_ZONE("Gen.Griffin");
KVCache& kv_cache = kv_caches[0];
hwy::ThreadPool& pool = activations.env.Pool();
namespace hn = hwy::HWY_NAMESPACE;
using D = hn::ScalableTag<float>;
static constexpr size_t kModelDim = TConfig::kModelDim;
static constexpr size_t kConv1dWidth = TConfig::kConv1dWidth;
static constexpr size_t kHeads = TConfig::kHeads;
// X / Y linear layers.
for (size_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
float* HWY_RESTRICT y = activations.griffin_y.Batch(batch_idx);
float* HWY_RESTRICT x = activations.griffin_x.Batch(batch_idx);
TwoMatVecAdd<kModelDim, kModelDim>(
layer_weights->griffin.linear_x_w, layer_weights->griffin.linear_y_w, 0,
activations.pre_att_rms_out.Batch(batch_idx),
/*add0=*/layer_weights->griffin.linear_x_biases.data_scale1(),
/*add1=*/layer_weights->griffin.linear_y_biases.data_scale1(),
/*out0=*/x, /*out1=*/y, pool);
Gelu(y, kModelDim);
}
// 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.Batch(batch_idx);
HWY_FULL(float) df;
HWY_DASSERT(kModelDim % hn::Lanes(df) == 0);
const size_t layer_offset = layer * kModelDim * (kConv1dWidth - 1);
// cache[i] = input at time t-i.
float* HWY_RESTRICT cache[HWY_MAX(kConv1dWidth, 1)];
cache[0] = x;
for (size_t i = 1; i < kConv1dWidth; i++) {
cache[i] =
kv_cache.conv1d_cache.get() + layer_offset +
((pos + kConv1dWidth - 1 - i) % (kConv1dWidth - 1)) * kModelDim;
}
for (size_t i = 0; i < kModelDim; i += hn::Lanes(df)) {
auto xv = hn::Load(df, x + i);
auto accum0 =
hn::Load(df, layer_weights->griffin.conv_biases.data_scale1() + i);
auto accum1 = hn::Zero(df);
static_assert(kConv1dWidth % 2 == 0, "Conv width must be even");
for (size_t l = 0; 2 * l < kConv1dWidth; l++) {
auto wv0 = hn::Load(df, layer_weights->griffin.conv_w.data_scale1() +
(kConv1dWidth - 1 - 2 * l) * kModelDim + i);
auto wv1 = hn::Load(df, layer_weights->griffin.conv_w.data_scale1() +
(kConv1dWidth - 2 - 2 * l) * kModelDim + 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(kConv1dWidth, 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.Batch(batch_idx);
float* HWY_RESTRICT x = activations.griffin_x.Batch(batch_idx);
float* HWY_RESTRICT gate_x = activations.griffin_gate_x.Batch(batch_idx);
float* HWY_RESTRICT a = activations.griffin_multiplier.Batch(batch_idx);
float* HWY_RESTRICT rnn_state =
kv_cache.rglru_cache.get() + layer * kModelDim;
pool.Run(0, kHeads, [&](const uint64_t head, size_t /*thread*/) HWY_ATTR {
constexpr size_t kHeadDim = kModelDim / kHeads;
constexpr size_t kMatrixSize = kHeadDim * kHeadDim;
size_t head_offset = head * kHeadDim;
TwoOfsMatVecAddLoop<kHeadDim, kHeadDim>(
layer_weights->griffin.gate_w, kMatrixSize * head,
kMatrixSize * (kHeads + head), x + head_offset,
/*add0=*/layer_weights->griffin.gate_biases.data_scale1() +
head_offset,
/*add1=*/layer_weights->griffin.gate_biases.data_scale1() +
kModelDim + 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.data_scale1() + 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.Batch(batch_idx);
float* out_ptr = activations.att_sums.Batch(batch_idx);
MatVecAdd<kModelDim, kModelDim>(
layer_weights->griffin.linear_out_w, 0, x,
layer_weights->griffin.linear_out_biases.data_scale1(), out_ptr, pool);
}
}
// Wrapper class; holds arguments in member variables to shorten call sites.
template <class TConfig>
class GemmaAttention {
static constexpr size_t kCacheLayerSize = CacheLayerSize<TConfig>()();
static constexpr size_t kCachePosSize = CachePosSize<TConfig>()();
static constexpr size_t kHeads = TConfig::kHeads;
static constexpr size_t kKVHeads = TConfig::kKVHeads;
static constexpr size_t kModelDim = TConfig::kModelDim;
static constexpr size_t kQKVDim = TConfig::kQKVDim;
static constexpr size_t kQStride = Activations::QStride<TConfig>();
static constexpr size_t kSeqLen = TConfig::kSeqLen;
static constexpr bool kIsMHA = Activations::IsMHA<TConfig>();
// The attention window usually starts at 0 unless `pos` is larger than
// the attention window size, then it is `pos` - window_size + 1.
static HWY_INLINE size_t StartPos(size_t pos, size_t layer) {
const size_t att_window_size = TConfig::kAttentionWindowSizes[layer];
return pos - std::min(att_window_size - 1, pos);
}
template <typename T>
HWY_INLINE void PositionalEncodingQK(const T* qk, size_t pos, size_t layer,
const float mul, T* qk_out) {
const float* inv_timescale = activations_.inv_timescale.Const();
// PostQKType::Rope
(void)layer;
if (TConfig::kUseHalfRope) {
hwy::CopyBytes(qk, qk_out, kQKVDim * sizeof(*qk));
Rope(qk_out, kQKVDim / 2, inv_timescale, pos);
MulByConst(mul, qk_out, kQKVDim);
} else {
RopeAndMulBy(mul, qk, kQKVDim, inv_timescale, pos, qk_out);
}
}
// Fills activations.q and computes KV. For kIsMHA, 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");
// For the computation of Q, K, and V, it is useful to remember that
// qkv_einsum_w has shape [(kHeads + kKVHeads * 2), kKQVDim, kModelDim]
// and kQStride = kQKVDim * (kIsMHA ? 3 : 1);
const auto pre_att_rms_out =
ConstMat(activations_.pre_att_rms_out.All(), kModelDim);
MatMul</*kAdd=*/false>(
num_interleaved, pre_att_rms_out,
ConstMat(layer_weights_.qkv_einsum_w.data(), kModelDim),
layer_weights_.qkv_einsum_w.scale(), /*add=*/nullptr, activations_.env,
MutableMat(activations_.q.All(), kHeads * kQStride));
if constexpr (kIsMHA) {
static_assert(TConfig::kInterleaveQKV, "MHA implies interleaved");
// Multi-Head Attention a.k.a. "use_qkv_einsum" computed QKV already.
} else {
// Single query and no wraparound means we can use a matmul and write
// directly into the KV cache with a stride of kCachePosSize.
if (num_queries_ == 1 &&
queries_pos_[0] + num_tokens_ <= div_seq_len_.GetDivisor()) {
const size_t kv_ofs =
queries_pos_[0] * kCachePosSize + layer_ * kCacheLayerSize;
// KV structure is [k, v, k, v, ....] = kKVHeads pairs of (k, v).
float* HWY_RESTRICT kv = kv_caches_[0].kv_cache.get() + kv_ofs;
MatMul</*kAdd=*/false>(
num_tokens_, pre_att_rms_out,
ConstMat(layer_weights_.qkv_einsum_w.data(), kModelDim, kModelDim,
kHeads * kQKVDim * kModelDim),
layer_weights_.qkv_einsum_w.scale(), /*add=*/nullptr,
activations_.env,
MutableMat(kv, kKVHeads * 2 * kQKVDim, kCachePosSize));
} 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.Batch(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 * kCachePosSize + layer_ * kCacheLayerSize;
float* HWY_RESTRICT kv = kv_cache.kv_cache.get() + kv_offset;
// KV structure is [k, v, k, v, ....] = kKVHeads pairs of (k, v).
MatVec<kKVHeads * 2 * kQKVDim, kModelDim>(
layer_weights_.qkv_einsum_w, kHeads * kQKVDim * kModelDim, x, kv,
pool_);
}
}
}
// Apply positional encodings for K (and copy KV to cache if MHA).
pool_.Run(
0, kKVHeads * num_interleaved,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % kKVHeads;
const size_t interleaved_idx = task / kKVHeads;
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 * kCachePosSize +
layer_ * kCacheLayerSize +
head * kQKVDim * 2;
KVCache& kv_cache = kv_caches_[query_idx];
float* HWY_RESTRICT kv = kv_cache.kv_cache.get() + kv_offset;
const float* HWY_RESTRICT mha_kv =
activations_.q.Batch(interleaved_idx) + head * kQStride + kQKVDim;
// Copy from `q` if MHA, or apply in-place.
PositionalEncodingQK(kIsMHA ? mha_kv : kv, pos, layer_, 1.0f, kv);
// If MHA, also copy V into KVCache.
if (kIsMHA) {
hwy::CopyBytes(mha_kv + kQKVDim, kv + kQKVDim,
kQKVDim * sizeof(*kv));
}
});
}
// Computes Q.K scores, which are "logits" (or scores) stored to head_att.
HWY_INLINE void QDotK(const size_t start_pos, const size_t last_pos,
const size_t head_offset, const float* HWY_RESTRICT q,
const KVCache& kv_cache, float* HWY_RESTRICT head_att) {
if (HWY_LIKELY(last_pos < kSeqLen)) {
// Slightly faster: no wraparound.
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const size_t kv_offset =
pos * kCachePosSize + layer_ * kCacheLayerSize + head_offset;
const float* HWY_RESTRICT k = &kv_cache.kv_cache[kv_offset];
const float score = Dot(q, k, kQKVDim);
head_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 size_t kv_offset =
cache_pos * kCachePosSize + layer_ * kCacheLayerSize + head_offset;
const float* HWY_RESTRICT k = &kv_cache.kv_cache[kv_offset];
const float score = Dot(q, k, kQKVDim);
head_att[pos % kSeqLen] = score;
}
}
}
// Accumulates the sum of v (from `kv_cache`) * probability (`head_att`) into
// `att_out`. Equivalent in gemma/modules.py:
// encoded = jnp.einsum('BTNS,BSNH->BTNH', probs, value_proj)
static HWY_INLINE void WeightedSumV(
const size_t start_pos, const size_t last_pos,
const float* HWY_RESTRICT head_att, const size_t layer,
const size_t head_offset, const hwy::Divisor& div_seq_len,
const KVCache& kv_cache, float* HWY_RESTRICT att_out) {
hwy::ZeroBytes(att_out, kQKVDim * sizeof(*att_out));
if (HWY_LIKELY(last_pos < kSeqLen)) {
// Slightly faster: no wraparound.
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const size_t kv_offset =
pos * kCachePosSize + layer * kCacheLayerSize + head_offset;
const float* HWY_RESTRICT v =
kv_cache.kv_cache.get() + kv_offset + kQKVDim;
MulByConstAndAdd(head_att[pos], v, att_out, kQKVDim);
}
} else {
for (size_t pos = start_pos; pos <= last_pos; ++pos) {
const size_t cache_pos = div_seq_len.Remainder(pos);
const size_t kv_offset =
cache_pos * kCachePosSize + layer * kCacheLayerSize + head_offset;
const float* HWY_RESTRICT v =
kv_cache.kv_cache.get() + kv_offset + kQKVDim;
MulByConstAndAdd(head_att[pos % kSeqLen], v, att_out, kQKVDim);
}
}
}
HWY_NOINLINE void DotSoftmaxWeightedSum(const size_t num_interleaved) {
PROFILER_ZONE("Gen.Attention.DotSoftmax");
GEMMA_CONSTEXPR_SQRT float kQueryScale = ChooseQueryScale<TConfig>();
// A "head group" in the context of GQA refers to a collection of query
// heads that share the same key and value heads.
static_assert((kHeads % kKVHeads) == 0,
"query heads must be a multiple of key-value heads");
constexpr size_t kHeadGroups = kHeads / kKVHeads;
// For each head (token, query), compute Q.K, softmax, and weighted V.
pool_.Run(0, kHeads * num_interleaved,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % kHeads;
const size_t interleaved_idx = task / kHeads;
const size_t query_idx = interleaved_idx % num_queries_;
const size_t batch_idx = interleaved_idx / num_queries_;
const size_t head_offset = (head / kHeadGroups) * kQKVDim * 2;
KVCache& kv_cache = kv_caches_[query_idx];
float* HWY_RESTRICT q =
activations_.q.Batch(interleaved_idx) + head * kQStride;
// Apply rope and scaling to Q.
const size_t pos = queries_pos_[query_idx] + batch_idx;
PositionalEncodingQK(q, pos, layer_, kQueryScale, q);
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;
}
float* HWY_RESTRICT head_att =
activations_.att.Batch(interleaved_idx) + head * kSeqLen;
QDotK(start_pos, last_pos, head_offset, q, kv_cache, head_att);
// SoftMax with optional SoftCap yields "probabilities" in
// head_att.
const size_t head_att_len = std::min(last_pos + 1, kSeqLen);
MaybeLogitsSoftCap(TConfig::kAttCap, head_att, head_att_len);
Softmax(head_att, head_att_len);
float* HWY_RESTRICT att_out =
activations_.att_out.Batch(interleaved_idx) +
head * kQKVDim;
WeightedSumV(start_pos, last_pos, head_att, layer_, head_offset,
div_seq_len_, kv_cache, att_out);
});
}
// Sums encoded (`att_out`) over num_heads (`kHeads`) and head_dim (`kQKVDim`)
// into output (`layer_out`).
HWY_NOINLINE void SumHeads(const size_t num_interleaved) {
PROFILER_ZONE("Gen.Attention.SumHeads");
constexpr bool kAdd = TConfig::kSoftmaxAttnOutputBiases;
const float* bias =
kAdd ? layer_weights_.attention_output_biases.data_scale1() : nullptr;
// att_weights and att_out are concatenated heads, each of length kQKVDim.
// Thus the [num_interleaved, kModelDim] matmul output is the sum over
// heads. Compare gemma/modules.py:
// attn_output = self.attn_vec_einsum('BTNH,NHD->BTD', encoded)
MatMul<kAdd>(
num_interleaved, ConstMat(activations_.att_out.All(), kHeads * kQKVDim),
ConstMat(layer_weights_.att_weights.data(), kHeads * kQKVDim),
layer_weights_.att_weights.scale(), bias, activations_.env,
MutableMat(activations_.att_sums.All(), kModelDim));
}
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 CompressedLayer<TConfig>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches)
: queries_pos_(queries_pos),
queries_prefix_end_(queries_prefix_end),
num_queries_(queries_pos.size()),
num_tokens_(num_tokens),
layer_(layer),
activations_(activations),
layer_weights_(*layer_weights),
div_seq_len_(div_seq_len),
kv_caches_(kv_caches),
pool_(activations.env.Pool()) {
HWY_DASSERT(num_queries_ <= kv_caches_.size());
}
// 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 CompressedLayer<TConfig>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches)
: queries_pos_(queries_pos),
queries_prefix_end_vec_(queries_pos.size(), 0),
queries_prefix_end_(queries_prefix_end_vec_.data(),
queries_prefix_end_vec_.size()),
num_queries_(queries_pos.size()),
num_tokens_(num_tokens),
layer_(layer),
activations_(activations),
layer_weights_(*layer_weights),
div_seq_len_(div_seq_len),
kv_caches_(kv_caches),
pool_(activations.env.Pool()) {
HWY_DASSERT(num_queries_ <= kv_caches_.size());
}
// 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(num_interleaved);
}
private:
const QueriesPos& queries_pos_;
const std::vector<size_t> queries_prefix_end_vec_;
const QueriesPos queries_prefix_end_;
const size_t num_queries_;
const size_t num_tokens_;
const size_t layer_;
Activations& activations_;
const CompressedLayer<TConfig>& layer_weights_;
const hwy::Divisor& div_seq_len_;
const KVCaches& kv_caches_;
hwy::ThreadPool& pool_;
};
template <class TConfig>
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 CompressedLayer<TConfig>* layer_weights,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches) {
if (type == LayerAttentionType::kGemma) {
GemmaAttention<TConfig>(queries_pos, queries_prefix_end, num_tokens, layer,
activations, layer_weights, div_seq_len,
kv_caches)();
} else {
// Only reached if the model is Griffin. `if constexpr` prevents generating
// this code for non-Griffin models.
if constexpr (TConfig::kGriffinLayers > 0) {
HWY_ASSERT(queries_pos.size() == 1);
GriffinRecurrent<TConfig>(queries_pos[0], num_tokens, layer, 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 <class TConfig>
class VitAttention {
static constexpr size_t kHeads = TConfig::kHeads;
static constexpr size_t kKVHeads = TConfig::kKVHeads;
static constexpr size_t kModelDim = TConfig::kModelDim;
static constexpr size_t kQKVDim = TConfig::kQKVDim;
static constexpr size_t kQStride = 3 * kQKVDim;
static constexpr size_t kSeqLen = TConfig::kSeqLen;
// Computes Q, K, V for all heads, stored in activations_.q.
HWY_NOINLINE void ComputeQKV() {
PROFILER_ZONE("Gen.VitAttention.QKV");
const auto y =
ConstMat(activations_.pre_att_rms_out.All(), kModelDim);
auto& qkv = activations_.q;
HWY_ASSERT(qkv.BatchSize() == num_tokens_);
HWY_ASSERT(qkv.Len() == kHeads * kQStride);
MatMul</*kAdd=*/true>(
num_tokens_, y,
ConstMat(layer_weights_.vit.qkv_einsum_w.data_scale1(), kModelDim),
/*scale=*/1.0f, layer_weights_.vit.qkv_einsum_b.data_scale1(),
activations_.env, MutableMat(qkv.All(), qkv.Len()));
}
HWY_NOINLINE void DotSoftmaxWeightedSum() {
GEMMA_CONSTEXPR_SQRT float kQueryScale =
1.0f / Sqrt(static_cast<float>(TConfig::kQKVDim));
PROFILER_ZONE("Gen.VitAttention.DotSoftmax");
// A "head group" in the context of GQA refers to a collection of query
// heads that share the same key and value heads.
static_assert(kHeads == kKVHeads, "Vit expects MHA");
// Compute Q.K, softmax, and weighted V.
pool_.Run(0, kHeads * num_tokens_,
[&](uint64_t task, size_t /*thread*/) HWY_ATTR {
const size_t head = task % kHeads;
const size_t token = task / kHeads;
// Compute Q.K scores, which are "logits" stored in head_att.
float* HWY_RESTRICT q =
activations_.q.Batch(token) + head * kQStride;
MulByConst(kQueryScale, q, kQKVDim);
float* HWY_RESTRICT head_att =
activations_.att.Batch(token) + head * kSeqLen;
for (size_t i = 0; i < kSeqLen; ++i) {
float* HWY_RESTRICT k =
activations_.q.Batch(i) + head * kQStride + kQKVDim;
head_att[i] = Dot(q, k, kQKVDim); // score = q.k
}
// SoftMax yields "probabilities" in head_att.
Softmax(head_att, kSeqLen);
// Compute weighted sum of v into att_out.
float* HWY_RESTRICT att_out =
activations_.att_out.Batch(token) + head * kQKVDim;
hwy::ZeroBytes(att_out, kQKVDim * sizeof(*att_out));
for (size_t i = 0; i < kSeqLen; ++i) {
float* HWY_RESTRICT v =
activations_.q.Batch(i) + head * kQStride + 2 * kQKVDim;
MulByConstAndAdd(head_att[i], v, att_out, kQKVDim);
}
});
}
// Sums encoded (`att_out`) over num_heads (`kHeads`) and head_dim (`kQKVDim`)
// into output (`att_sums`).
HWY_NOINLINE void SumHeads() {
PROFILER_ZONE("Gen.VitAttention.SumHeads");
auto* bias = layer_weights_.vit.attn_out_b.data_scale1();
auto att_out = ConstMat(activations_.att_out.All(), kHeads * kQKVDim);
auto att_weights = ConstMat(layer_weights_.vit.attn_out_w.data_scale1(),
kHeads * kQKVDim);
auto att_sums = MutableMat(activations_.att_sums.All(), kModelDim);
// att_weights and att_out are concatenated heads, each of length kQKVDim.
// Thus the [num_tokens_, kModelDim] matmul output is the sum over heads.
MatMul</*kAdd=*/true>(num_tokens_, att_out, att_weights, /*scale=*/1.0f,
bias, activations_.env, att_sums);
}
public:
VitAttention(size_t num_tokens, size_t layer, Activations& activations,
const CompressedLayer<TConfig>* layer_weights)
: num_tokens_(num_tokens),
layer_(layer),
activations_(activations),
layer_weights_(*layer_weights),
pool_(activations.env.Pool()) {}
HWY_INLINE void operator()() {
ComputeQKV();
DotSoftmaxWeightedSum();
SumHeads();
}
private:
const size_t num_tokens_;
const size_t layer_;
Activations& activations_;
const CompressedLayer<TConfig>& layer_weights_;
hwy::ThreadPool& pool_;
};
template <class TConfig, typename T>
HWY_NOINLINE void Activation(T* HWY_RESTRICT c1, 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));
});
}
template <class TConfig>
HWY_NOINLINE void FFW(Activations& activations, size_t num_interleaved,
const CompressedLayer<TConfig>* layer_weights) {
PROFILER_ZONE("Gen.FFW");
constexpr size_t kModelDim = TConfig::kModelDim;
constexpr size_t kFFHiddenDim = TConfig::kFFHiddenDim;
constexpr bool kAddBias = TConfig::kFFBiases;
constexpr bool kIsVit = TConfig::kLayerConfig[0] == LayerAttentionType::kVit;
using WeightType =
hwy::If<kIsVit,
typename CompressedLayer<TConfig>::WeightF32OrBF16,
typename CompressedLayer<TConfig>::Weight>;
HWY_DASSERT(num_interleaved <= activations.bf_pre_ffw_rms_out.BatchSize());
// Define slightly more readable names for the weights and activations.
const auto x = ConstMat(activations.bf_pre_ffw_rms_out.All(), kModelDim);
Mat<const WeightType> w1;
const float* bias1 = nullptr;
Mat<const WeightType> w2;
const float* bias2 = nullptr;
float scale = 1.0f;
Mat<const WeightType> w_output;
const float* output_bias = nullptr;
float output_scale = 1.0f;
auto hidden_activations = MutableMat(activations.C1.All(), kFFHiddenDim);
auto multiplier = MutableMat(activations.C2.All(), kFFHiddenDim);
auto ffw_out = MutableMat(activations.ffw_out.All(), kModelDim);
// For some of the weights and activations, it depends on the config where to
// get them from or whether to use them at all.
if constexpr (kAddBias && !kIsVit) {
bias1 = layer_weights->ffw_gating_biases.data_scale1();
bias2 = bias1 + kFFHiddenDim;
output_bias = layer_weights->ffw_output_biases.data_scale1();
}
if constexpr (!kIsVit) {
w1 = ConstMat(layer_weights->gating_einsum_w.data(), kModelDim);
w2 = ConstMat(layer_weights->gating_einsum_w.data(), kModelDim, kModelDim,
kModelDim * kFFHiddenDim);
scale = layer_weights->gating_einsum_w.scale();
w_output = ConstMat(layer_weights->linear_w.data(), kFFHiddenDim);
output_scale = layer_weights->linear_w.scale();
} else {
w1 = ConstMat(layer_weights->vit.linear_0_w.data_scale1(), kModelDim);
bias1 = layer_weights->vit.linear_0_b.data_scale1();
multiplier.ptr = nullptr;
w_output =
ConstMat(layer_weights->vit.linear_1_w.data_scale1(), kFFHiddenDim);
output_bias = layer_weights->vit.linear_1_b.data_scale1();
}
// Compute the hidden layer activations.
MatMul<kAddBias>(num_interleaved, x, w1, scale, bias1, activations.env,
hidden_activations);
if constexpr (!kIsVit) {
MatMul<kAddBias>(num_interleaved, x, w2, scale, bias2, activations.env,
multiplier);
}
// Activation (Gelu) and maybe multiply by gate. Store activations in act.
Activation<TConfig>(hidden_activations.ptr, multiplier.ptr,
kFFHiddenDim * num_interleaved);
// Hidden layer -> output layer.
MatMul<kAddBias>(num_interleaved, ConstMat(hidden_activations), w_output,
output_scale, 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.
template <class TConfig>
HWY_NOINLINE void EmbedToken(int token, size_t batch_idx, size_t pos,
size_t pos_in_prompt,
const CompressedWeights<TConfig>& weights,
RowVectorBatch<float>& x,
const ImageTokens* image_tokens) {
// Image tokens just need to be copied.
if (image_tokens != nullptr && pos_in_prompt < image_tokens->BatchSize()) {
hwy::CopyBytes(image_tokens->Batch(pos_in_prompt), x.Batch(batch_idx),
x.Len() * sizeof(x.Const()[0]));
return;
}
constexpr size_t kModelDim = TConfig::kModelDim;
constexpr size_t kVocabSize = TConfig::kVocabSize;
GEMMA_CONSTEXPR_EMBSCALING const float kEmbScaling =
EmbeddingScaling<TConfig>();
HWY_DASSERT(token >= 0);
HWY_DASSERT(token < kVocabSize);
const hn::ScalableTag<float> df;
DecompressAndZeroPad(
df,
MakeSpan(weights.embedder_input_embedding.data(), kVocabSize * kModelDim),
token * kModelDim, x.Batch(batch_idx), kModelDim);
MulByConst(kEmbScaling * weights.embedder_input_embedding.scale(),
x.Batch(batch_idx), kModelDim);
if constexpr (TConfig::kAbsolutePE) {
AddAbsolutePositionalEmbeddings(x.Batch(batch_idx), kModelDim, pos);
}
}
template <class TConfig, typename T>
HWY_NOINLINE void ResidualConnection(
size_t num_interleaved, T* HWY_RESTRICT other, T* HWY_RESTRICT x,
const CompressedLayer<TConfig>* layer_weights, bool is_attention) {
constexpr size_t kModelDim = TConfig::kModelDim;
// ResidualType::Add
AddFromBatched(num_interleaved, other, x, kModelDim);
}
template <class TConfig, typename WeightT, typename InOutT>
void PostNorm(size_t num_interleaved, const WeightT& weights, InOutT* inout) {
if (TConfig::kPostNorm == PostNormType::Scale) {
RMSNormInplaceBatched(num_interleaved, weights.data_scale1(), inout,
TConfig::kModelDim);
}
}
template <class TConfig>
HWY_NOINLINE void TransformerLayer(
const QueriesPos& queries_pos, const QueriesPos& queries_prefix_end,
size_t num_tokens, size_t layer,
const CompressedLayer<TConfig>* layer_weights, Activations& activations,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches) {
constexpr size_t kModelDim = TConfig::kModelDim;
const size_t num_interleaved = num_tokens * queries_pos.size();
auto type = TConfig::kLayerConfig[layer];
size_t layer_of_type =
NumLayersOfTypeBefore(TConfig::kLayerConfig, type, layer);
RMSNormBatched(num_interleaved, activations.x.All(),
layer_weights->pre_attention_norm_scale.data_scale1(),
activations.pre_att_rms_out.All(), kModelDim);
Attention<TConfig>(type, queries_pos, queries_prefix_end, num_tokens,
layer_of_type, activations, layer_weights, div_seq_len,
kv_caches);
PostNorm<TConfig>(num_interleaved, layer_weights->post_attention_norm_scale,
activations.att_sums.All());
ResidualConnection<TConfig>(num_interleaved, activations.att_sums.All(),
activations.x.All(), layer_weights,
/*is_attention=*/true);
RMSNormBatched(num_interleaved, activations.x.All(),
layer_weights->pre_ffw_norm_scale.data_scale1(),
activations.bf_pre_ffw_rms_out.All(), kModelDim);
FFW<TConfig>(activations, num_interleaved, layer_weights);
PostNorm<TConfig>(num_interleaved, layer_weights->post_ffw_norm_scale,
activations.ffw_out.All());
ResidualConnection<TConfig>(num_interleaved, activations.ffw_out.All(),
activations.x.All(), 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 mergig this with TransformerLayer.
template <class TConfig>
HWY_NOINLINE void VitTransformerLayer(
size_t num_tokens, size_t layer,
const CompressedLayer<TConfig>* layer_weights, Activations& activations) {
constexpr size_t kModelDim = TConfig::kModelDim;
auto type = TConfig::kLayerConfig[layer];
HWY_ASSERT(type == LayerAttentionType::kVit);
auto& x = activations.x;
HWY_ASSERT(x.BatchSize() == num_tokens);
HWY_ASSERT(x.Len() == kModelDim);
// y = nn.LayerNorm()(x)
// y ~ pre_att_rms_out
LayerNormBatched(num_tokens, x.All(),
layer_weights->vit.layer_norm_0_scale.data_scale1(),
layer_weights->vit.layer_norm_0_bias.data_scale1(),
activations.pre_att_rms_out.All(), kModelDim);
// y = out["sa"] = nn.MultiHeadDotProductAttention(...)(y, y)
// y ~ att_sums
VitAttention<TConfig>(num_tokens, layer, activations, layer_weights)();
// x = out["+sa"] = x + y
AddFromBatched(num_tokens, activations.att_sums.All(), x.All(), kModelDim);
// y = nn.LayerNorm()(x)
// y ~ bf_pre_ffw_rms_out
LayerNormBatched(num_tokens, x.All(),
layer_weights->vit.layer_norm_1_scale.data_scale1(),
layer_weights->vit.layer_norm_1_bias.data_scale1(),
activations.bf_pre_ffw_rms_out.All(), kModelDim);
// y = out["mlp"] = MlpBlock(...)(y)
// y ~ ffw_out
FFW<TConfig>(activations, num_tokens, layer_weights);
// x = out["+mlp"] = x + y
AddFromBatched(num_tokens, activations.ffw_out.All(), x.All(), kModelDim);
}
// 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 <class TConfig>
HWY_NOINLINE void Prefill(
const QueriesPromptTokens& queries_prompt,
const QueriesMutablePos& queries_pos, const QueriesPos& queries_prefix_end,
const size_t query_idx_start, const CompressedWeights<TConfig>& 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_ASSERT(queries_pos.size() == num_queries);
HWY_ASSERT(queries_prefix_end.size() == num_queries);
HWY_ASSERT(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.BatchSize();
// 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?
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(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);
// Fill activations.x (much faster than TransformerLayer).
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];
EmbedToken<TConfig>(token, ti, pos, pos_in_prompt, weights,
activations.x, runtime_config.image_tokens);
}
// Transformer with one batch of tokens from a single query.
for (size_t layer = 0; layer < TConfig::kLayers; ++layer) {
const auto* layer_weights = weights.GetLayer(layer);
TransformerLayer<TConfig>(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 <class TConfig>
HWY_NOINLINE void EmbedImagePatches(const Image& image,
const CompressedWeights<TConfig>& weights,
Activations& activations) {
static constexpr size_t kModelDim = TConfig::VitConfig::kModelDim;
static constexpr size_t kPatchWidth = TConfig::VitConfig::kPatchWidth;
static constexpr size_t kSeqLen = TConfig::VitConfig::kSeqLen;
constexpr size_t kPatchSize = kPatchWidth * kPatchWidth * 3;
HWY_ASSERT(weights.vit_img_embedding_kernel.NumElements() ==
kPatchSize * kModelDim);
HWY_ASSERT(activations.x.Len() == kModelDim);
std::vector<hwy::AlignedFreeUniquePtr<float[]>> image_patches(kSeqLen);
for (size_t i = 0; i < kSeqLen; ++i) {
image_patches[i] = hwy::AllocateAligned<float>(kPatchSize);
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:
// RowVectorBatch<float> image_patches(kSeqLen, kPatchSize);
// [Get patches]
// MatMul</*kAdd=*/true>(
// kVitSeqLen, ConstMat(image_patches.All(), kPatchSize),
// ConstMat(weights.vit_img_embedding_kernel.data_scale1(), kPatchSize),
// /*scale=*/1.0f, weights.vit_img_embedding_bias.data_scale1(),
// activations.env, MutableMat(activations.x.All(), 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 < kSeqLen; ++i) {
MatVecAdd<kModelDim, kPatchSize>(
weights.vit_img_embedding_kernel, 0, image_patches[i].get(),
weights.vit_img_embedding_bias.data_scale1(), activations.x.Batch(i),
activations.env.Pools().Outer());
}
// Add position embeddings.
AddFrom(weights.vit_img_pos_embedding.data_scale1(), activations.x.All(),
kSeqLen * kModelDim);
}
// Prefills the image tokens with the ViT encoder.
template <class TConfig>
HWY_NOINLINE void PrefillVit(const CompressedWeights<TConfig>& weights,
const RuntimeConfig& runtime_config,
const Image& image, ImageTokens& image_tokens,
Activations& activations) {
PROFILER_ZONE("Gen.PrefillVit");
const size_t num_tokens = TConfig::VitConfig::kSeqLen;
const size_t kVitModelDim = TConfig::VitConfig::kModelDim;
HWY_ASSERT(num_tokens == activations.x.BatchSize());
// Embed the image patches.
EmbedImagePatches<TConfig>(image, weights, activations);
// Go through all layers.
for (size_t layer = 0; layer < TConfig::VitConfig::kLayers; ++layer) {
const auto* layer_weights = weights.GetVitLayer(layer);
VitTransformerLayer<typename TConfig::VitConfig>(
num_tokens, layer, layer_weights, activations);
}
// Final Layernorm.
LayerNormBatched(num_tokens, activations.x.All(),
weights.vit_encoder_norm_scale.data_scale1(),
weights.vit_encoder_norm_bias.data_scale1(),
activations.x.All(), kVitModelDim);
// Apply head embedding into image_tokens of size of the LLM kModelDim.
MatMul</*kAdd=*/true>(
num_tokens, ConstMat(activations.x.All(), kVitModelDim),
ConstMat(weights.vit_img_head_kernel.data_scale1(), kVitModelDim),
/*scale=*/1.0f, weights.vit_img_head_bias.data_scale1(), activations.env,
MutableMat(image_tokens.All(), TConfig::kModelDim));
}
// 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 <class TConfig>
HWY_NOINLINE void Transformer(
const QueriesToken& queries_token, const QueriesMutablePos& queries_pos,
const QueriesPos& queries_prefix_end,
const CompressedWeights<TConfig>& weights, Activations& activations,
const hwy::Divisor& div_seq_len, const KVCaches& kv_caches,
const LayersOutputFunc& layers_output,
const ActivationsObserverFunc& activations_observer) {
constexpr size_t kModelDim = TConfig::kModelDim;
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);
}
}
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
EmbedToken<TConfig>(queries_token[query_idx], query_idx,
queries_pos[query_idx], /*pos_in_prompt=*/0, weights,
activations.x, /*image_tokens=*/nullptr);
}
for (size_t layer = 0; layer < TConfig::kLayers; ++layer) {
const CompressedLayer<TConfig>* layer_weights = weights.GetLayer(layer);
TransformerLayer<TConfig>(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(num_queries, weights.final_norm_scale.data_scale1(),
activations.x.All(), kModelDim);
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;
}
}
template <class TConfig>
void RangeChecks(size_t& max_tokens, size_t& max_generated_tokens,
size_t& prompt_size) {
if (!TConfig::kUseLocalAttention) {
if (max_tokens > TConfig::kSeqLen) {
fprintf(stderr, "WARNING: max_tokens %zu > kSeqLen %d, truncating.\n",
max_tokens, TConfig::kSeqLen);
max_tokens = static_cast<size_t>(TConfig::kSeqLen);
}
}
if (max_generated_tokens > max_tokens) {
fprintf(stderr,
"WARNING: max_generated_tokens %zu > max_tokens %zu, truncating.\n",
max_generated_tokens, max_tokens);
max_generated_tokens = max_tokens - 1;
}
if (!TConfig::kUseLocalAttention) {
if (prompt_size + max_generated_tokens > max_tokens) {
fprintf(stderr,
"WARNING: prompt_size %zu + max_generated_tokens %zu > "
"max_tokens %zu, truncating to ",
prompt_size, max_generated_tokens, max_tokens);
prompt_size = std::min(prompt_size, max_tokens - max_generated_tokens);
fprintf(stderr, "%zu\n", prompt_size);
}
}
HWY_ASSERT(prompt_size > 0);
}
// 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)
: runtime_config_(runtime_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) ||
token == runtime_config_.eos_id) {
is_eos_.Set(query_idx);
return true;
}
return false;
}
private:
const RuntimeConfig& runtime_config_;
hwy::BitSet4096<> is_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 <class TConfig>
void GenerateT(const ByteStorageT& weights_u8, 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) {
constexpr size_t kModelDim = TConfig::kModelDim;
constexpr size_t kVocabSize = TConfig::kVocabSize;
const CompressedWeights<TConfig>& weights =
*reinterpret_cast<const CompressedWeights<TConfig>*>(weights_u8.get());
// 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());
const size_t num_queries = queries_prompt.size();
HWY_ASSERT(num_queries <= 4096); // TokenStreamer uses BitSet4096.
HWY_ASSERT(num_queries <= activations.x.BatchSize());
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_tokens = runtime_config.max_tokens;
size_t max_generated_tokens = runtime_config.max_generated_tokens;
RangeChecks<TConfig>(max_tokens, max_generated_tokens, max_prompt_size);
for (size_t pos : queries_pos_copy) {
if (pos >= max_tokens) {
fprintf(stderr, "Warning: pos %zu >= max_tokens %zu, aborting.\n", pos,
max_tokens);
return;
}
}
// If no sample_func is provided, we use top-k sampling.
const SampleFunc sample_token =
runtime_config.sample_func
? runtime_config.sample_func
: [&](const float* logits, size_t vocab_size) -> int {
return SampleTopK<TConfig::kTopK>(logits, vocab_size, *runtime_config.gen,
runtime_config.temperature,
runtime_config.accept_token);
};
// Prefill stops before min_prompt_size - 1 because the last prompt token is
// the first input token for generation.
const double 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.BatchSize();
Activations prefill_activations;
if (use_prefill_activations) {
prefill_activations.Allocate<TConfig>(runtime_config.prefill_tbatch_size,
activations.env.Pools());
}
Prefill<TConfig>(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, prefill_start);
// 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);
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 double gen_start = hwy::platform::Now();
for (size_t gen = 0; gen < HWY_MIN(max_tokens, max_generated_tokens); ++gen) {
// Decode generates one token per query and increments queries_mutable_pos.
Transformer<TConfig>(
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.
bool all_queries_eos = true;
PROFILER_ZONE("Gen.Embedding");
// Compute logits from last layer activations.
MatMul</*kAdd=*/false>(
num_queries, ConstMat(activations.x.All(), kModelDim),
ConstMat(weights.embedder_input_embedding.data(), kModelDim),
weights.embedder_input_embedding.scale(), /*add=*/nullptr,
activations.env, MutableMat(activations.logits.All(), kVocabSize));
for (size_t query_idx = 0; query_idx < num_queries; ++query_idx) {
float* HWY_RESTRICT logits = activations.logits.Batch(query_idx);
MaybeLogitsSoftCap(TConfig::kFinalCap, logits, kVocabSize);
Softmax(logits, kVocabSize);
const int token = sample_token(logits, kVocabSize);
timing_info.NotifyGenerated(prefill_start, gen_start);
const bool is_eos =
token_streamer(query_idx_start + query_idx,
queries_mutable_pos[query_idx], token, logits[token]);
all_queries_eos &= is_eos;
gen_tokens[query_idx] = is_eos ? runtime_config.eos_id : token;
}
if (all_queries_eos) break;
} // foreach token to generate
timing_info.NotifyGenerateDone(gen_start);
}
template <class TConfig>
void GenerateSingleT(const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config,
const PromptTokens& prompt, size_t pos, size_t prefix_end,
KVCache& kv_cache, PerClusterPools& pools,
TimingInfo& timing_info) {
constexpr size_t kNumQueries = 1;
const size_t qbatch_start = 0;
// TODO: move into Gemma?
Activations activations;
activations.Allocate<TConfig>(kNumQueries, pools);
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<TConfig>(weights_u8, activations, runtime_config, queries_prompt,
queries_pos, queries_prefix_end, qbatch_start, kv_caches,
timing_info);
}
template <class TConfig>
void GenerateBatchT(const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt,
const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end,
const KVCaches& kv_caches, PerClusterPools& pools,
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.
const size_t max_qbatch_size =
(TConfig::kGriffinLayers > 0) ? 1 : runtime_config.decode_qbatch_size;
Activations activations;
activations.Allocate<TConfig>(max_qbatch_size, pools);
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);
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<TConfig>(weights_u8, activations, runtime_config, qbatch_prompts,
qbatch_pos, qbatch_prefix_end, qbatch_start, qbatch_kv,
timing_info);
}
}
template <class TConfig>
void GenerateImageTokensT(const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config,
const Image& image, ImageTokens& image_tokens,
PerClusterPools& pools) {
if constexpr (TConfig::VitConfig::kLayers == 0) {
return;
} else {
Activations prefill_activations;
RuntimeConfig prefill_runtime_config = runtime_config;
prefill_runtime_config.prefill_tbatch_size = TConfig::VitConfig::kSeqLen;
prefill_activations.Allocate<typename TConfig::VitConfig>(
prefill_runtime_config.prefill_tbatch_size, pools);
// Weights are for the full PaliGemma model, not just the ViT part.
const CompressedWeights<TConfig>& weights =
*reinterpret_cast<const CompressedWeights<TConfig>*>(weights_u8.get());
PrefillVit<TConfig>(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_CONFIG, which is for function overloading.
void GenerateSingle( // NOLINT(misc-definitions-in-headers)
GEMMA_CONFIG, const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config, const PromptTokens& prompt, size_t pos,
size_t prefix_end, KVCache& kv_cache, PerClusterPools& pools,
TimingInfo& timing_info) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateSingleT<GEMMA_CONFIG>)
(weights_u8, runtime_config, prompt, pos, prefix_end, kv_cache, pools,
timing_info);
}
void GenerateBatch( // NOLINT(misc-definitions-in-headers)
GEMMA_CONFIG, const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config,
const QueriesPromptTokens& queries_prompt, const QueriesPos& queries_pos,
const QueriesPos& queries_prefix_end, const KVCaches& kv_caches,
PerClusterPools& pools, TimingInfo& timing_info) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateBatchT<GEMMA_CONFIG>)
(weights_u8, runtime_config, queries_prompt, queries_pos, queries_prefix_end,
kv_caches, pools, timing_info);
}
void GenerateImageTokens( // NOLINT(misc-definitions-in-headers)
GEMMA_CONFIG, const ByteStorageT& weights_u8,
const RuntimeConfig& runtime_config, const Image& image,
ImageTokens& image_tokens, PerClusterPools& pools) {
HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(GenerateImageTokensT<GEMMA_CONFIG>)
(weights_u8, runtime_config, image, image_tokens, pools);
}
#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