Skip to content

[fix] Normalization compute step for non-nvidia hardware #84

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mikepapadim merged 2 commits into from
Dec 11, 2025
Merged

[fix] Normalization compute step for non-nvidia hardware #84

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b7bf484
[fix] Norm step for non-nvidia hardware
mikepapadim Dec 11, 2025
7b830ea
Update src/main/java/org/beehive/gpullama3/tornadovm/layers/type/q8_0…
mikepapadim Dec 11, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/Phi3FP16FFNLayers.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,6 +234,12 @@ TaskGraph setupSinglePhi3FFNLayer(Phi3TornadoWeights weights, int layerIndex) {
phi3Config.rmsNormEps(), // epsilon
phi3State.localSize); // local memory size

if (shouldUseFinalNormalization()) {
unifiedLayer.task("attn_rms_finalize",
TransformerComputeKernelsLayered::reductionFinalNormalization,
context, state.temp, config.dim(), config.rmsNormEps());
}

unifiedLayer.task("attn_rms_qkv_projection", Phi3Kernels::fusedRmsNormQKVMatmulDirect,
context, phi3State.wrapX, // input
phi3State.wrapQ, // output Q
Expand Down
9 changes: 9 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/Qwen2FP16FFNLayers.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -271,6 +271,15 @@ TaskGraph setupSingleQwen2FFNLayer(Qwen2TornadoWeights weights, int layerIndex)
config.rmsNormEps(), // epsilon
qwen2State.localSize); // local memory size

if (shouldUseFinalNormalization()) {
unifiedLayer.task("attn_rms_finalize",
TransformerComputeKernelsLayered::reductionFinalNormalization,
context,
state.temp,
config.dim(),
config.rmsNormEps());
}

// Fused RMS Apply + QKV Projection
unifiedLayer.task("attn_rms_qkv_projection",
Qwen3Kernels::fusedRmsNormQKVMatmul,
Expand Down
9 changes: 9 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/Qwen3FP16FFNLayers.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -264,6 +264,15 @@ TaskGraph setupSingleQwen3FFNLayer(Qwen3TornadoWeights weights, int layerIndex)
qwen3Config.rmsNormEps(), // epsilon
qwen3State.localSize); // local memory size

if (shouldUseFinalNormalization()) {
unifiedLayer.task("attn_rms_finalize",
TransformerComputeKernelsLayered::reductionFinalNormalization,
context,
state.temp,
config.dim(),
config.rmsNormEps());
}

// Fused RMS Apply + QKV Projection
unifiedLayer.task("attn_rms_qkv_projection",
Qwen3Kernels::fusedRmsNormQKVMatmul,
Expand Down
9 changes: 9 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/layers/type/q8_0/Phi3Q8_0FFNLayers.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -236,6 +236,15 @@ TaskGraph setupSinglePhi3Q8_0FFNLayer(Phi3TornadoWeights weights, int layerIndex
phi3Config.rmsNormEps(), // epsilon
phi3State.localSize); // local memory size

if (shouldUseFinalNormalization()) {
unifiedLayer.task("attn_rms_finalize",
TransformerComputeKernelsLayered::reductionFinalNormalization,
context,
state.temp,
config.dim(),
config.rmsNormEps());
}

// Fused: RMS apply + Q8 QKV matmul + direct Q/K/V split
unifiedLayer.task("attn_rms_qkv_projection_q8",
TransformerComputeKernelsLayered::fusedRmsNormQKVMatmulQ8,
Expand Down
9 changes: 9 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/layers/type/q8_0/Qwen3Q8_0FFNLayers.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,6 +190,15 @@ TaskGraph setupSingleQwen3FFNLayer(Qwen3TornadoWeights weights, int layerIndex)
config.rmsNormEps(), // epsilon
qwen3State.localSize); // local memory size

if (shouldUseFinalNormalization()) {
unifiedLayer.task("attn_rms_finalize",
TransformerComputeKernelsLayered::reductionFinalNormalization,
context,
state.temp,
config.dim(),
config.rmsNormEps());
}

// Fused RMS Apply + QKV Projection
unifiedLayer.task("attn_rms_qkv_projection",
Qwen3Kernels::fusedRmsNormQKVMatmulQ8_0,
Expand Down