Conversation
added 2 commits
August 1, 2026 22:47
Adapt FlagTensor to Iluvatar BI-V150 (CoreX SDK 4.4.0, CUDA-compatible): - Detect placeholder libcutensor.so shipped by the CoreX SDK (dlopens but lacks CUTENSOR_COMPUTE_DESC_* symbols) and mark cuTensor unavailable instead of crashing at import; full NVIDIA installs are unaffected since every probed symbol is already required there. - New _iluvatar vendor backend: PyTorch-native ops as the vendor baseline (no cuTensor equivalent exists on Iluvatar), vendor tolerances (1e-3 benchmark-verify floor), elementwise heuristics, ixsmi device query. - Route devices named 'Iluvatar ...' to the iluvatar backend in DeviceDetector. - Fix a multi-process race in unary kernel codegen: write _gen_*.py atomically and skip rewriting unchanged content, so parallel run_tests.py workers can no longer read half-written files. Verified on 4x BI-V150: python tools/run_tests.py --stages all --gpus 0,1,2,3 -> accuracy 36/36, performance 36/36.
Contraction 0.51x -> 1.66x and ContractionTrinary 0.65x -> 2.21x vs the Iluvatar-native baseline; suite geomean 1.82x -> 1.95x. All changes are gated behind the iluvatar vendor; other vendors keep byte-identical autotune spaces and launch paths. - Append 4 BI-V150-tuned configs (small 32x32 tiles, few stages) to the GETT autotune space when the iluvatar backend is active. - Break down the 'batched shapes' regression: the CoreX Triton fork's JITFunction.run dispatch costs ~0.15 ms per call vs ~0.01 ms of GPU time, which dominates host-timed paths. After the first autotuner-driven call, freeze the winning config into a direct CompiledKernel runner (~0.012 ms per call). CUDA-graph replay was measured but rejected (~0.06 ms GPU-side launch overhead). - Time the production two-step GETT path in the ContractionTrinary benchmark for fp32 as well (the op never dispatches the fused kernel: _supports_fused_triton_trinary() is always False).
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.
天数适配初版代码合入