Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Annotating parameters with tl.pointer_type results in KeyError #4690

Open
carmocca opened this issue Sep 10, 2024 · 1 comment
Open

Annotating parameters with tl.pointer_type results in KeyError #4690

carmocca opened this issue Sep 10, 2024 · 1 comment

Comments

@carmocca
Copy link

Annotating parameters with foo: tl.pointer_type results in the following stacktrace

my_model.py:90: in forward
    my_kernel[lambda meta: (triton.cdiv(numel, meta["BLOCK_SIZE"]),)](
../../../../.local/lib/python3.10/site-packages/triton/runtime/jit.py:345: in <lambda>
    return lambda *args, **kwargs: self.run(grid=grid, warmup=False, *args, **kwargs)
../../../../.local/lib/python3.10/site-packages/triton/runtime/autotuner.py:156: in run
    timings = {config: self._bench(*args, config=config, **kwargs) for config in pruned_configs}
../../../../.local/lib/python3.10/site-packages/triton/runtime/autotuner.py:156: in <dictcomp>
    timings = {config: self._bench(*args, config=config, **kwargs) for config in pruned_configs}
../../../../.local/lib/python3.10/site-packages/triton/runtime/autotuner.py:133: in _bench
    return do_bench(kernel_call, warmup=self.num_warmups, rep=self.num_reps, quantiles=(0.5, 0.2, 0.8))
../../../../.local/lib/python3.10/site-packages/triton/testing.py:103: in do_bench
    fn()
../../../../.local/lib/python3.10/site-packages/triton/runtime/autotuner.py:114: in kernel_call
    self.fn.run(
../../../../.local/lib/python3.10/site-packages/triton/runtime/jit.py:662: in run
    kernel = self.compile(
../../../../.local/lib/python3.10/site-packages/triton/compiler/compiler.py:276: in compile
    module = src.make_ir(options, codegen_fns, context)
../../../../.local/lib/python3.10/site-packages/triton/compiler/compiler.py:113: in make_ir
    return ast_to_ttir(self.fn, self, context=context, options=options, codegen_fns=codegen_fns)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
name = 'ier_type'
    def str_to_ty(name):
        if name[0] == "*":
            name = name[1:]
            if name[0] == "k":
                name = name[1:]
                ty = str_to_ty(name)
                return const_pointer_type(ty)
            ty = str_to_ty(name)
            return pointer_type(ty)
        tys = {
            "fp8e4nv": float8e4nv,
            "fp8e4b8": float8e4b8,
            "fp8e5": float8e5,
            "fp8e5b16": float8e5b16,
            "fp8e4b15": float8e4b15,
            "fp16": float16,
            "bf16": bfloat16,
            "fp32": float32,
            "fp64": float64,
            "i1": int1,
            "i8": int8,
            "i16": int16,
            "i32": int32,
            "i64": int64,
            "u1": int1,
            "u8": uint8,
            "u16": uint16,
            "u32": uint32,
            "u64": uint64,
            "B": int1,
        }
>       return tys[name]
E       KeyError: 'ier_type'
../../../../.local/lib/python3.10/site-packages/triton/language/__init__.py:[284]

The fix is simple: to remove these annotations.

But this used to work in previous versions.

Environment:

triton                    3.0.0
pytorch-triton            2.2.0+e28a256d71
torch                     2.4.1+cu118

Thanks!

@iclementine
Copy link

I do some debugging and find that the bug lies in the generated binder.

class JITFunction(KernelInterface[T]):
    def run(self, *args, grid, warmup, **kwargs):
        if self.binder is None:
            self.create_binder()
        ...
        bound_args, sig_and_spec, constexpr_vals, non_constexpr_vals, excess_kwargs = self.binder(*args, **kwargs)

The sig_and_spec returned from the binder is already wrong.

I made a jit function with the following signature

@triton.jit
def unary_pointwise_kernel(X: tl.pointer_type, O: tl.pointer_type, n: int, BLOCK_N: tl.constexpr):

the sig_and_spec is ('ier_type', 'ier_type', 'i32', 'D', 'D', 'D').

Finally I found that the bug is in

class KernelParam:
    ...
    @cached_property
    def annotation_type(self):
        annotation = self.annotation
        for ty1, ty2 in [("uint", 'u'), ("int", 'i')]:
            width = annotation[annotation.find(ty1) + len(ty1):]
            if width and ty1 in annotation:
                return f"{ty2}{width}"
        if annotation == "bool":
            return "u1"
        return ""

here annotatin is pointer_type (tl. is stripped beforehand), and int is in pointer_type. So width is er_type, the return value is ier_type.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants