This PR adds an autoload mechanism for out-of-tree backends, simplifies
TVM's Python library loading, and removes `TVMError` in favor of native
Python errors.
## Autoload out-of-tree backends
Out-of-tree packages can register an autoload callable under the
`tvm.backends` entry-point group (mirroring torch's device-backend
autoload). At `import tvm` startup each entry point is discovered and
its callable invoked once, after the core runtime and the `tvm`
namespace are fully initialized, so an extension can register
ops/targets/funcs or load extra libraries.
```toml
[project.entry-points."tvm.backends"]
tvm_foo = "tvm_foo:_autoload"
```
A failing extension is caught and surfaced via `warnings.warn` so it
cannot break `import tvm`. Autoload can be disabled with
`TVM_DEVICE_BACKEND_AUTOLOAD=0`.
## Simplify library loading
The library-loading path in `base.py` is consolidated around a single
`_LOADED_LIBS` dict (basename to ctypes handle) so downstream and
autoloaded extensions can skip already-loaded libraries; the per-backend
runtime DSO list is folded into `load_backend_libs`. Accumulated cruft
is removed: the Python-3.9 check, the readline shim, the `_FFI_MODE`
ctypes check, the `base.__version__` re-export, and `py_str` (call sites
inline `.decode("utf-8")`).
## Remove TVMError in favor of native Python errors
`TVMError` added a layer atop `RuntimeError` that downstream code had to
import and learn. It is removed; the registered FFI error kinds
(`InternalError`, `RPCError`, `OpError`, `DiagnosticError`,
`ScheduleError`) now subclass `RuntimeError` directly while staying
registered, so the FFI keeps throwing the right kinds. All `TVMError`
imports, `except`/`raise`/`isinstance` uses, and
`pytest.raises(tvm.TVMError)` sites move to the `RuntimeError` builtin.
* [FFI][REFACTOR] Establish tvm_ffi as a standalone python module
This PR establishes tvm_ffi as a standalone python module.
The ffi is structured as a minimal pip module that can be
directly install by path or url.
examples/get_started provided a minimal example.
This is a major change as we are decoupling tvm_ffi as a
separate package, users need to install tvm_ffi separately.
Thanks to its minimal dependency, tvm_ffi can be easily installed
even just from the source by pip install ./ffi
This change would enable future improvement for library plugins
to have lightweight dependencies by just working on top of
the tvm_ffi, while the main compiler toolchain and runtime
can be layered on top.
* [FFI] Improve traceback setups
This PR improves traceback related setups
This PR modernizes the FFI foundation of the project and introduce
a new minimal and lightweight module [tvm ffi](https://github.com/apache/tvm/tree/refactor-s3/ffi)
based on our lessons in the past few years. It implements a modern
version of the [Unified Packed and Object RFC](https://github.com/apache/tvm-rfcs/blob/main/rfcs/0097-unify-packed-and-object.md)
that unifies the packed function call and object systems.
Summary of the change:
- A dedicated clean Any/AnyView that can store strong and weak
references of items
- Function(previously PackedFunc) system built on top of the Any/AnyView
- A minimal C API that backs the overall calls. We are stabilizing the
API with a goal to bring clean, stable FFI conventions for both compiled
and registered code
- A rewrite of core python binding and generated code based on the module
- Update existing code and test cases to the new module
- Latest dlpack support
The new module brings many benefits thanks to the cleaner design,
to name a few:
- Any can support both POD types(int) and object types.
- Containers (e.g. Array) can now also contain Any value, e.g. now
`Array<int>` is supported, no need for boxed types
- Error handling now upgrades to object-based, allowing cleaner
traceback across languages
- Map now preserves insertion orders
- Path toward isolated stabilize minimum core ABI/API foundation module
- Type traits based design that cleanly defines how values interact
with Any system
- Automatic conversion of different types based on traits if needed
Because FFI upgrade is at heart of the project, the change touches every
component of the system. Importantly, this is an upgrade of the ABI so the
change is not backward compatible. The code compiled under the old
FFI won't work under the new one. We did provide example ABI translation
(e.g. LegacyTVMArgValueToFFIAny) functions for compatibility.
The PR tries to leave files in their old places while creating redirections.
The goal is to have the first milestone landed and infrastructure in place,
so we can do further refactors to complete features and cleanup legacy code
as trackable PRs. As of now, python binding and compiled code are under the
new convention while RPC and some other bindings still relies on legacy ABI
translation. We will work on upgrades in the coming PRs, including areas such
as reflection, phasing out legacy redirections etc.
* [TIR][Analysis] Implemented tir.analysis.is_pure_function
This commit introduces two related utilities,
`tir.analysis.is_pure_function` and `tir.analysis.assert_pure_function`.
In contrast to the existing `tvm::tir::SideEffect`, which checks for
side effects on a for a `PrimExpr`, `is_pure_function` checks for side
effects for the function as a whole.
* [Transform] Implement relax.transform.ComputePrimValue
Prior to this commit, while expressions of type `DataType::Int(64)`
could be computed in the `relax.transform.VMShapeLower`, expressions
of any other type could not. This commit introduces
`relax.transform.ComputePrimValue`, which produces `PrimFunc`
subroutines to compute `PrimExpr` values of any dtype.
This functionality will allow boolean values to be computed based on
the symbolic values known at runtime.
* [Relax] Allow R.Prim('bool') in relax::If and assert_op
Prior to this commit, the condition used for `relax::If` node and the
`"relax.assert_op"` operator was required to be a scalar tensor. This
made it difficult to alter behavior based on a runtime shape
parameter. For example, delegating to a vectorized implementation
based on a whether a tensor shape is divisible by the vector size.
This commit adds support for expressions of type `R.Prim('bool')` as
the conditional for `relax::If` and `"relax.assert_op"`, to allow
these use cases.
* Lint fix
By using RPC server in NPU board, at some time a compiled model will hang the NPU, because of the buggy operator libraries of NPU toolchain, so we must to use the session_timeout to ensure the board resource can be released by the hang jobs.
Currently the handling of session timeout error in RPC server is not good, it just kill the server loop sub process, then in the destructor of class `RPCEndpoint` will send the code of `kShutdown` to the RPC client, but the RPC client expect receive the code of `kReturn` or `kException`, so users will see the error message that like the one reported in https://github.com/apache/tvm/issues/15151, this error report will make users very confused and don't know what's happened.
When using tuning to search a good schedule for operators, we only want to ignore the RPC session timeout error that indicate the schedule generated is an illegal one, but other error reported by the RPC server may help us find the potential bug of our tool chain built on top of TVM, so the RPC session timeout error should be split to a standalone TVM error class.
This PR implemented these requirements by sending the RPC session timeout error message as a PRC server exception to the RPC client before kill the server loop sub process.
This PR introduces some cleanups on the stacktraces in TVM's error message by removing the meaningless frames, for example, C++ template dispatch with `std::enable_if`.
* Rename references to 0.0.0.0 to localhost. Also change references to 127.0.0.1 to localhost so that all references are consistent. 0.0.0.0 is not the same as localhost.
* Add code from livestream with JK
* Fix errors parsing ResNet
* Parse metadata section efficiently and do most of plumbing to resolve metadata section references.
* WIP
* Change meta reference to an operator
* Meta references now work
* MetaReference expansion now works
* Start working on source map and move diagnostic context
* Convert tokenizer and parser to use new machinery
* Kill to_json
* Fix comment in type_infer.cc
* Remove old parser
* Rename parser tests and remove old ones
* Record span end information
* Convert to using spans everywhere
* Add span fields back to all Relay constructors
* Start passing spans
* Pass spans around visitors
* Format
* Fix
* Fix
* disable reference lint from string helpers
* Fix tokenizer
* Fix issue with empty metadata section
* Document new span fields and small tweaks
* Formatting
* Add span doc fields
* Add format tweak
* Improve errors and fix the semantic version tags in Prelude
* Update gradient.rly
* Clean up broken spans
* Clean up parser tests and turn on previously skipped tests
* Update errors to handle skipped cases
* Tweak
* Tweak
* Format
* Fix some minor issues with ADT tests
* Format
* Fix path
* WIP
* WIP
* Fix ir_text_printer
* format
* Formatted
* More formatting
* Repair test cases
* Fix CI
* Retrigger CI
* Update dmlc-core which was mistakenly overriden
* [REFACTOR][RPC][PROCOTOL-CHANGE] Modularize the RPC infra.
This PR refactors the RPC protocol to make it more modularized.
- RPCSession: represent a set of features that need to be implemented
- RPCEndPont: End point that forwards the RPCSession requests over a communication channel.
- RPCModule: Exposes an RPCSession as an rpc device in the TVM Runtime API.
In the new design, the local machine is presented as a special case of RPCSession.
The remote is just another client session that calls into RPCEndPoint.
The RPC communication path is as follows.
```
client -> ClientSession -> EndPoint[client@n0]
-> networking[between n0 <=> n1]
-> EndPoint[server@n1] -> LocalSession[@n1]
```
Because of the new modular design, we can now chain more sessions together.
For example, we can now run the following proxy setup (testcase in test_runtime_rpc.test_session_constructor).
```
client -> ClientSession -> Endpoint[client@n0]
-> networking[between n0 <=> n1]
-> Endpoint[server@n1] -> ClientSession -> Endpoint[client@n1]
-> networking[between n1 <=> n2]
-> Endpoint[server@n2] -> LocalSession[@n2]
```
We can also implement other types of Sessions.
As an example, We introduced a PopenSession that communicates with
the another process via a pipe.
We also add more comments about the internal of the RPC.
The communication protocol is simplfied using a similar convention as PackedFunc.
This allows us to further reduce the amount of special remote syscalls.
Due to the major improvement and simplification, we are making a non-compatible update to the RPC protocol.
It means that the client and server needs to be upgraded to together in order for it to function correctly.
This PR also introduces a versioning mechanism to the current RPC procotol,
so that future upgrade will be produce more user friendly with error messages.
* Address review comments
* Remove ld library path
This PR enables the copy on write optimizations passes:
- Enable COW for IRModule both TIR and relay passes.
- Enabled COW for PrimFunc in TIR passes.
Need more thoughts into whether/how to enable COW
for relay::Function, due to some function passes depend
on the presence of IRModule for context information,
and the std::move of the related function to nullptr
might affect the related behavior.
* [REFACTOR][NODE][RUNTIME] Move Node to the new Object protocol.
This PR removes the original node system, and make node as a subclass of Object.
This is a major refactor towards a better unified runtime object system.
List of changes in the refactor:
- We now hide data_ field, use Downcast explicitly to get a sub-class object.
- Removed the node system FFI in python.
- Removed the node C API, instead use PackedFunc for list and get attrs.
- Change relay::Op::set_attr_type_key(attr_key_name) to relay::Op::set_attr_type<AttrType>().
- This change was necessary because of the new Object registration mechanism.
- Subsequent changes to the op registrations
- The change revealed a few previous problems that is now fixed.
- Patched up a few missing node type registration.
- Now we will raise an error if we register object that is not registered.
- The original node.h and container.h are kept in the same location.
- Calling convention: kObjectHandle now equals the old kNodeHandle, kNodeHandle is removed.
- IRFunctor now dispatches on ObjectRef.
- Update to the new type checking API: is_type, derived_from are replaced by IsInstance.
- Removed .hash member function, instead use C++ convention hasher functors.
* Address review comments
Tianqi Chen