Develop

[YWHyuk]

Changelog — develop → master

TOGSim (simulator)

• Memory backend: updated to Ramulator 2.1.
• Config format: Configuration files have migrated from JSON to YAML format.
• Stats & robustness: Clearer DRAM bandwidth reporting, safer idle-stat handling, fixes for local/remote memory stats.
• Scheduling: Internal graph API cleanup (non-breaking, no user-facing API changes).Trace files support comments; improved CLI help.

Compiler & runtime (PyTorchSim / MLIR)

• PyTorch version: 2.1 → 2.8 (PyTorch version update #196)
• Operators: SDPA can now be routed to a dedicated NPU kernel via torch.nn.attention.sdpa_kernel([SDPBackend.FLASH_ATTENTION]) context manager; TopK, Bitonic sort, Cat added. ([BUG]Support for repeat_interleave operation to enable Grouped Query Attention (GQA) #198)
• CNNs: MobileNet CI and 1×1 spatial conv as linear; baseline group convolution decomposition + tests. ([BUG] Cannot schedule MobileNet-SSDLite model #205)
• Dtypes / codegen: Fixed float16 codegen in MLIR templates; worked around gem5 lmul8 widening issue by avoiding the problematic vector-width in codegen.
• TOGSim session: Run kernels under with TOGSimulator(config_path=...): so config and simulator lifecycle are scoped to the block.
• Multi-tenant launch: Call torch.npu.launch_model(opt_fn, *args, stream_index=..., timestamp=..., **kwargs) inside that block.
• Cleanup: Removed legacy scheduler code; standardized on the TOGSimulator-oriented API.

Device (OpenReg / NPU)

• Device API: Use torch.device("npu") (and torch.device("npu:0"), etc.) like any built-in device type — no extra package import beyond import torch; the NPU backend registers with PyTorch's device system.
• Eager mode: CPU fallback is applied automatically when graph compilation is not available.

⚠️ Breaking Changes

• Config format migration: Configuration files must be converted from JSON to YAML format. Existing .json config files are no longer supported.
• Multi-tenant API redesign: The scheduler-based multi-tenant launch pattern has been replaced. The old API required manual Scheduler instantiation, Request object construction, and a while not scheduler.is_finished(): loop. The new API uses a with TOGSimulator(config_path=...): context and torch.npu.launch_model(..., stream_index=..., timestamp=...) calls directly. See test_scheduler.py for the updated usage pattern.

CI, tests, experiments

• Added or tightened tests for DeepSeek, YOLOv5, MobileNet; CI image updated for PyTorch 2.8.

Other

• Misc. codegen, indexing, and matmul-related bugfixes and small refactors.