$ the-wire · showcase
TorchTitan pipeline and CP sharding land, ExecuTorch quantization finally holds
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
10 people shipped this
TorchTitan's Kimi K3 text decoder gains pipeline parallelism, ExecuTorch makes per-channel weight quantization survive the full AoT-to-kernel path, and pkg-config consumers stop needing a source build.
[Kimi K3] Pipeline Parallelism support (text LLM side): PP Virtual Stage Cache Adapter for Attention Residual pytorch/torchtitan
The Kimi K3 text decoder now takes pipeline_parallel_degree > 1; parallelize.py previously rejected it outright. Because core's pipeline_llm carries one hidden state per hop and cannot express Block Attention Residuals, a PP Virtual Stage Cache Adapter accumulates every earlier block's residual and runs output_res_proj and output_res_norm only on the stage owning lm_head.
Support per-channel weights end to end (#22227) pytorch/executorch
Per-channel weight quantization used to pass prepare_pt2/convert_pt2 and then quietly drop back to float, because fusion bailed on the per-channel dequant and convs fell out to cadence.fully_connected; on HiFi op_quantized_linear_out would requantize every channel with channel 0's scale. The whole path now works end to end, AoT and kernels, with granularity carried through.
[CP] Decouple input and metadata sharding with configurable CP load balancer pytorch/torchtitan
Context-parallel load balancing is now one explicit per-batch token permutation shared by model inputs and attention metadata, with head-tail and PTRR generating the permutations and None reserved for Ulysses. ContextParallelPartitioner and private PyTorch CP sharding helpers are replaced by TorchTitan's shard_inputs(), which applies the permutation and shards declared tensors via spmd.shard().
Make core MoE collectives explicit (#4878) pytorch/torchtitan
All core MoE collectives are now explicit, with automatic src to dst redistribution removed and collectives placed inside the appropriate remat regions. SharedExpertRowParallelLinear is added along the way.
Ship a pkg-config file for the runtime in the Linux and macOS wheels pytorch/executorch
The Linux and macOS wheels now ship lib/pkgconfig/executorch.pc, so build systems that read pkg-config, Meson among them, no longer need a full source build just for that file. Paths resolve from the file's own location, so it works wherever pip installed the wheel, and it emits the same compile definitions as the CMake package's runtime.
Revert "[symm_mem]: Add per-PG stream serialization for ops (#195947)" pytorch/pytorch
The quieter tail: pytorch/pytorch reverted the symm_mem per-PG stream serialization after test_stream_serialization_put_wait_two_streams failed deterministically on linux.dgx.b200.8 with 64 of 64 mismatched on rank 3, and landed an inductor fix that records device-TMA emission in store_reduction() so tma_min_block_sizes survives in-kernel descriptor formatting.
