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

Factor out SumHeads 

PiperOrigin-RevId: 842138081
This commit is contained in:
Martin Stolle 2025-12-09 02:22:46 -08:00 committed by Copybara-Service
parent 1014ae9e2a
commit 14a9ecf21d
2 changed files with 20 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
gemma/attention.cc
View File

@ -43,6 +43,7 @@
// After highway.h // After highway.h
#include "compression/compress-inl.h" #include "compression/compress-inl.h"
#include "gemma/flash_attention.h" #include "gemma/flash_attention.h"
#include "gemma/gemma-inl.h"
#include "ops/ops-inl.h" #include "ops/ops-inl.h"
HWY_BEFORE_NAMESPACE(); HWY_BEFORE_NAMESPACE();
@ -319,25 +320,6 @@ static HWY_INLINE void ComputeQKV(size_t num_tokens, const size_t layer_idx,
}); });
} }
// Sums encoded (`att_out`) over num_heads (`layer_config.heads`) and
// head_dim (`qkv_dim`) into output (`layer_out`).
static HWY_INLINE void SumHeads(const LayerWeightsPtrs& layer,
AttentionActivationsPtrs& activations,
MatMulEnv& env) {
GCPP_ZONE(env.ctx, hwy::Profiler::GlobalIdx(), Zones::kGenAttentionSumHeads);
const LayerConfig& layer_config = layer.layer_config;
(void)layer_config; // For HWY_DASSERT
// 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);
CallMatMul(activations.att_out, layer.att_weights, /*add=*/nullptr, env,
activations.att_sums);
}
void GemmaAttention(size_t num_tokens, const size_t layer_idx, void GemmaAttention(size_t num_tokens, const size_t layer_idx,
const LayerWeightsPtrs& layer, const LayerWeightsPtrs& layer,
AttentionActivationsPtrs& activations, QBatch& qbatch, AttentionActivationsPtrs& activations, QBatch& qbatch,

19
gemma/gemma-inl.h
View File

@ -183,6 +183,25 @@ static inline void FFWNoVit(const LayerWeightsPtrs& layer,
CallMatMul(activations.C1, layer.linear_w, nullptr, env, activations.ffw_out); CallMatMul(activations.C1, layer.linear_w, nullptr, env, activations.ffw_out);
} }
// Sums encoded (`att_out`) over num_heads (`layer_config.heads`) and
// head_dim (`qkv_dim`) into output (`layer_out`).
static HWY_INLINE void SumHeads(const LayerWeightsPtrs& layer,
AttentionActivationsPtrs& activations,
MatMulEnv& env) {
GCPP_ZONE(env.ctx, hwy::Profiler::GlobalIdx(), Zones::kGenAttentionSumHeads);
const LayerConfig& layer_config = layer.layer_config;
(void)layer_config; // For HWY_DASSERT
// 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);
CallMatMul(activations.att_out, layer.att_weights, /*add=*/nullptr, env,
activations.att_sums);
}
// NOLINTNEXTLINE(google-readability-namespace-comments) // NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE } // namespace HWY_NAMESPACE
} // namespace gcpp } // namespace gcpp