Integrate Triton up to [50d803cd](https://github.com/openai/triton/co…

…mmits/50d803cdb4e68910ed663251100e168ea4d2519d

PiperOrigin-RevId: 673813747

jax_triton/triton_lib.py

@@ -391,7 +391,7 @@ def get_or_create_triton_kernel(
   if num_ctas > 1 and compute_capability < 90:
     raise ValueError("num_ctas > 1 unsupported before Hopper.")
 
-  signature = dict(enumerate(arg_dtypes))
+  signature = {fn.arg_names[i]: v for i, v in enumerate(arg_dtypes)}
   # TODO(sharadmv,zhangqiaorjc): handle differently aligned pointers
   # We assume that all arrays are aligned to 16 bytes, and Triton may use this
   # assumption, unless array args are include in the `do_not_specialize` list.
@@ -403,9 +403,9 @@ def get_or_create_triton_kernel(
     args_for_specialization_attr[i] = v
   specialization_attr = fn._get_config(*args_for_specialization_attr)  # pylint: disable=protected-access
 
-  constants = {fn.arg_names.index(k): v for k, v in metaparams.items()}
-  constants.update({i: None for i, _, v in scalar_args if v is None})
-  constants.update({i: 1 for i in specialization_attr.equal_to_1})
+  constants = {k: v for k, v in metaparams.items()}
+  constants.update({k: None for _, k, v in scalar_args if v is None})
+  constants.update({fn.arg_names[i]: 1 for i in specialization_attr.equal_to_1})
 
   # Cache key should contain any parameter that can affect the compiler output.
   cache_key = (