fix(compute): reuse dst GPU memory instead of allocating per call (#84)#85
Merged
fix(compute): reuse dst GPU memory instead of allocating per call (#84)#85
Conversation
GPU ops (gpuBinaryOp, gpuUnaryOp, gpuScalarOp, Transpose, MatMul, Sum) were allocating fresh device memory via pool.Alloc on every call even when a pre-sized dst tensor was provided, then swapping dst's storage to the new allocation. The old GPUStorage was orphaned and depended on Go's GC finalizer to call pool.Free. At large training shapes with hundreds of batches and ~20 ops per batch, orphaned allocations piled up faster than the GC could reclaim, causing unbounded GPU memory growth and OOM. Fix: add tryReuseDstPtr helper that checks if dst[0] already has a GPUStorage with sufficient capacity. If so, the kernel writes directly into the existing device pointer — no pool.Alloc, no orphaned storage, no GC pressure. When dst is nil or undersized, the existing alloc path is preserved unchanged. Applied to the six hot-path op families that cover PatchTST GPU training: - gpuBinaryOp (Add, Sub, Mul same-shape) - gpuUnaryOp (Exp, Log, Sin, Cos, Tanh, Sqrt) - gpuScalarOp (MulScalar, AddScalar, DivScalar) - Transpose (gpu_engine_memory.go) - MatMul standard float32 path (gpu_engine.go) - Sum/ReduceSum (gpu_kernels.go) Other ops (broadcast, Q4/Q8/BF16 matmul, fused kernels) continue using the existing alloc path and can be converted incrementally. Full ztensor test suite passes on CPU host. Closes #84 Refs zerfoo/zerfoo#373
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes #84. GPU ops were allocating fresh device memory via pool.Alloc on every call even when a pre-sized dst tensor was provided, orphaning old GPUStorage objects. At large training shapes, the GC couldn't keep pace with finalization, causing unbounded GPU memory growth → OOM.
Fix:
tryReuseDstPtrchecks if dst already has a GPUStorage with sufficient capacity. If so, the kernel writes directly into the existing device pointer — zero allocation, zero GC pressure. Applied to the six hot-path op families: gpuBinaryOp, gpuUnaryOp, gpuScalarOp, Transpose, MatMul, Sum.Validation on DGX GB10
Before fix (post-E85 commit
09a318c6):After fix:
Test plan
go build ./...cleango vet ./...clean (pre-existing unsafe.Pointer warnings only)go test ./...passes (all packages)Refs