additional changes from comments

plugins/framework/src/fms_acceleration/model_patcher.py

@@ -179,7 +179,7 @@ def __post_init__(self):
             self.forward is not None,
             self.forward_builder is not None,
             self.import_and_maybe_reload is not None,
-        ]) > 1:
+        ]) != 1:
             raise ValueError(
                 f"Rule '{self.rule_id}' must only have only one of forward, "
                 "foward builder, or import_and_maybe_reload, specified."
@@ -332,7 +332,7 @@ def _import_and_reload(model: torch.nn.Module):
 
         # If there are multiple reload targets,
         # ensure that their paths do not conflict as reloading same module might reset patches
-        if len(_with_reload)>1:
+        if len(_with_reload) > 1:
             # sort ascending target path length
             _with_reload = sorted(
                 _with_reload,

tox.ini

@@ -26,6 +26,10 @@ deps =
     -e {toxinidir}/plugins/framework # install the framework here as the flash attention deps requires torch
 passenv = * # will pass the parent env, otherwise there are too many envs e.g. TRANSFORMERS that need to be set
 setenv = 
+    # Need to be set in new versions of triton that don't allow for access to global variable in the JIT compile
+    # Subsequently, consider changing triton kernels to access global variables that are annotated as constexpr
+    # source: https://github.com/triton-lang/triton/blob/7b617bcc35c4cf06f61dd267fc049fe33b2851f9/python/triton/compiler/code_generator.py#L280
+    # Tracking this as an issue here # https://github.com/foundation-model-stack/fms-acceleration/issues/56
     TRITON_ALLOW_NON_CONSTEXPR_GLOBALS=1
 commands = 
     # need a version of fms-hf-tuning that has integrated the framework