Skip to content

[tornadovm] Replace the dummy forward pass in forceCopyInReadOnlyData() with TornadoExecutionPlan.transferToDevice() #148

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

Open
mikepapadim wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[tornadovm] Replace the dummy forward pass in forceCopyInReadOnlyData() with TornadoExecutionPlan.transferToDevice() #148

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
19 changes: 13 additions & 6 deletions src/main/java/org/beehive/gpullama3/tornadovm/TornadoVMMasterPlanBatchPrefillDecode.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,21 +78,28 @@ public TornadoExecutionPlan createExecutionPlan() {

// ── Initialisation ────────────────────────────────────────────────────────

/**
* Puts the read-only weights on the device. With CUDA graphs enabled the pass is also the
* capture, so it still runs every graph; otherwise it is a plain copy-in and no kernel runs.
*/
// @formatter:off
@Override
public void forceCopyInReadOnlyData() {
if (!CUDA_GRAPHS) {
executionPlan.transferToDevice();
return;
}

state.wrapX.clear();
state.positionHolder.init(0);
state.wrapXBatch.clear();
state.batchStartPosHolder.init(0);

for (int i = 0; i <= taskGraphLayout.logitsIdx(); i++) {
var g = executionPlan.withGraph(i)
.withGridScheduler(batchPrefillDecodeForwardPlan.getGridScheduler());
if (CUDA_GRAPHS) {
g.withCUDAGraph();
}
g.execute();
executionPlan.withGraph(i)
.withGridScheduler(batchPrefillDecodeForwardPlan.getGridScheduler())
.withCUDAGraph()
.execute();
}
}
// @formatter:on
Expand Down
20 changes: 13 additions & 7 deletions src/main/java/org/beehive/gpullama3/tornadovm/TornadoVMMasterPlanPrefillDecode.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,20 +86,26 @@ public TornadoExecutionPlan createExecutionPlan() {

// ── Initialisation ────────────────────────────────────────────────────────

/** Runs all graphs once to trigger FIRST_EXECUTION uploads and warm up CUDA graphs. */
/**
* Puts the read-only weights on the device. With CUDA graphs enabled the pass is also the
* capture, so it still runs every graph; otherwise it is a plain copy-in and no kernel runs.
*/
// @formatter:off
@Override
public void forceCopyInReadOnlyData() {
if (!CUDA_GRAPHS) {
executionPlan.transferToDevice();
return;
}

state.wrapX.clear();
state.positionHolder.init(0);

for (int i = 0; i <= taskGraphLayout.logitsIdx(); i++) {
var g = executionPlan.withGraph(i)
.withGridScheduler(prefillDecodeForwardPlan.getGridScheduler());
if (CUDA_GRAPHS) {
g.withCUDAGraph();
}
g.execute();
executionPlan.withGraph(i)
.withGridScheduler(prefillDecodeForwardPlan.getGridScheduler())
.withCUDAGraph()
.execute();
}
}
// @formatter:on
Expand Down
9 changes: 9 additions & 0 deletions src/main/java/org/beehive/gpullama3/tornadovm/TornadoVMMasterPlanSingleToken.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,9 +95,18 @@ public FloatArray tornadoVMForwardDecode(int position) {
}
// @formatter:on

/**
* Puts the read-only weights on the device. With CUDA graphs enabled the pass is also the
* capture, so it still runs every graph; otherwise it is a plain copy-in and no kernel runs.
*/
// @formatter:off
@Override
public void forceCopyInReadOnlyData() {
if (!CUDA_GRAPHS) {
executionPlan.transferToDevice();
return;
}

state.wrapX.clear();
state.positionHolder.init(0);

Expand Down
Loading