[frontend] Add graph partitioning, distributed TP execution, and RVV deployment for DeepSeek

[asdf1113]

Summary

This PR extends the frontend graph partitioning support and introduces distributed parallel execution for the DeepSeek tensor-parallel example.

On the frontend side, it enhances graph_driver so that a computation graph can be partitioned into multiple subgraphs at compile time. It also fixes and improves shape inference for several operators under graph partitioning.

On the example side, this PR adds a runnable DeepSeek Tensor Parallel example under examples/BuddyTensorParallel with distributed parallel execution support. In addition, this PR includes packaging and deployment support for RISC-V RVV targets, together with updated build and run documentation for both x86 and RVV workflows.

What’s included

• Extend graph_driver to support splitting a graph into multiple subgraphs
• Fix shape inference for ReshapeOp and ExpandOp under graph partitioning
• Improve output shape inference for FlashAttention and GQAAttention related operators
• Add a runnable DeepSeek tensor-parallel example under examples/BuddyTensorParallel
• Add runtime-level distributed execution support for the DeepSeek tensor-parallel example
• Add an MPICH-based multi-rank launch workflow for distributed tensor-parallel execution across multiple nodes
• Add RVV cross-compilation, packaging, and deployment support
• Update README and build/run instructions for both x86 and RVV usage

Notes

• This PR includes both frontend graph partitioning support and distributed parallel execution for the DeepSeek tensor-parallel example.
• The RVV workflow packages the executable, runtime dependencies, and rank-specific parameter files for deployment across multiple RISC-V machines.

[Copilot]

Pull request overview

This PR adds frontend graph partitioning support (splitting a compiled graph into multiple subgraphs) and introduces a DeepSeek-R1 TP=2 example that runs the resulting subgraphs sequentially, along with several shape-inference updates needed to keep partitioned graphs correct.

Changes:

• Add a SplitStrategy + extensive GraphDriver updates to split graphs into subgraphs and construct “main graphs” that call subgraphs.
• Fix/adjust shape inference behavior for reshape/expand and attention-related ops under partitioning.
• Add a new BuddyTensorParallel DeepSeek-R1 example (Python importer + C++ runner + CMake wiring) and extend MemRef utilities to support split/concat/add.

Reviewed changes

Copilot reviewed 14 out of 15 changed files in this pull request and generated 26 comments.

Show a summary per file

File	Description
frontend/Python/ops/tosa.py	Adjust reshape/expand and attention shape inference to work with partitioned graphs.
frontend/Python/ops/linalg.py	Adjust shape inference for several ops (pow/matmul/neg/cat).
frontend/Python/graph/type.py	Change TensorMeta to a dict-like container with property accessors.
frontend/Python/graph/transform/fuse_ops.py	Ensure fused attention op names are unique.
frontend/Python/graph/operation.py	Add _newshape support and new shape-splitting helper(s).
frontend/Python/graph/graph_driver.py	Major rewrite to implement subgraph splitting/strategy and main-graph construction.
frontend/Python/graph/graph.py	Track lowered output nodes and propagate _newshape into placeholder shapes during import.
frontend/Python/graph/init.py	Export SplitStrategy.
frontend/Interfaces/buddy/Core/Container.h	Add MemRef split/concat/add helpers and a constructor taking std::vector<size_t>.
examples/CMakeLists.txt	Add a build option/entry for BuddyTensorParallel examples.
examples/BuddyTensorParallel/import-deepseek-r1.py	New importer that partitions DeepSeek-R1 into subgraphs and emits MLIR + weight shards.
examples/BuddyTensorParallel/dis-main.cpp	New sequential TP example runner using split/concat/add MemRef helpers.
examples/BuddyTensorParallel/README.md	New documentation for building/running the example.
examples/BuddyTensorParallel/CMakeLists.txt	Build pipeline to generate MLIR and compile/link the example.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In construct_main_graph, the dependency wiring for CallOp inputs is broken: the check if node.name in self._graph.node_table is always true for any op in the original graph, so intermediate values produced by earlier subgraphs are incorrectly treated as external inputs. This should check whether the value is already available in main_graph.node_table (placeholders) and otherwise be sourced from an upstream CallOp result via _call_table.