Collapse _RuntimeCompiledFnInvoker.run,
_RuntimeForwardEpilogue.capture_orig_inputs,
increment_mutation_versions, and finalize into a single codegen'd
function with all branches resolved at compile time.

The generated function inlines:
- capture_orig_inputs: dict comprehension → baked {idx: args[idx]} literal
- increment_mutation_versions: conditional + generator → baked tuple
- compiled_invoker.run: trace_joint branch + detach indices inlined
- output arity validation: baked expected count
- split mutated inputs: baked slice index
- apply mutations / replay aliases: delegate to existing codegen'd functions
- dynamic dims: baked per-output dim sets
- grad_enabled_mutation: baked boolean

Generated code for inference (0 mutations, 1 alias, 1 input):

    def _runtime_wrapper(_compiled_fn_, _first_ctx_, _on_before_call_, args):
        orig_inputs = {0: args[0]}
        with _first_ctx_():
            grad_enabled = torch.is_grad_enabled()
            try:
                if grad_enabled: torch._C._set_grad_enabled(False)
                _on_before_call_()
                all_outs = _normalize_as_list_(_compiled_fn_(args))
            finally:
                if grad_enabled: torch._C._set_grad_enabled(True)
        del args
        if len(all_outs) != 1:
            raise AssertionError(...)
        fw_outs = all_outs
        ret_outs = _replay_aliases_(orig_inputs, fw_outs)
        return ret_outs

RuntimeWrapper orchestration step in isolation (us/call):

| Case | Before (method dispatch) | After (codegen) | Speedup |
|---|---|---|---|
| 0 alias, 0 mut, 5 args | 0.35 us | 0.17 us | 2.1x |
| 2 alias, 0 mut, 5 args | 0.41 us | 0.17 us | 2.5x |
| 0 alias, 2 mut, 5 args | 0.74 us | 0.25 us | 3.0x |
| 3 alias, 1 mut, 10 args | 0.79 us | 0.25 us | 3.2x |
| 5 alias, 3 mut, 20 args | 0.93 us | 0.32 us | 2.9x |

[ghstack-poisoned]

Collapse _RuntimeCompiledFnInvoker.run,
_RuntimeForwardEpilogue.capture_orig_inputs,
increment_mutation_versions, and finalize into a single codegen'd
function with all branches resolved at compile time.

The generated function inlines:
- capture_orig_inputs: dict comprehension → baked {idx: args[idx]} literal
- increment_mutation_versions: conditional + generator → baked tuple
- compiled_invoker.run: trace_joint branch + detach indices inlined
- output arity validation: baked expected count
- split mutated inputs: baked slice index
- apply mutations / replay aliases: delegate to existing codegen'd functions
- dynamic dims: baked per-output dim sets
- grad_enabled_mutation: baked boolean

Generated code for inference (0 mutations, 1 alias, 1 input):

    def _runtime_wrapper(_compiled_fn_, _first_ctx_, _on_before_call_, args):
        orig_inputs = {0: args[0]}
        with _first_ctx_():
            grad_enabled = torch.is_grad_enabled()
            try:
                if grad_enabled: torch._C._set_grad_enabled(False)
                _on_before_call_()
                all_outs = _normalize_as_list_(_compiled_fn_(args))
            finally:
                if grad_enabled: torch._C._set_grad_enabled(True)
        del args
        if len(all_outs) != 1:
            raise AssertionError(...)
        fw_outs = all_outs
        ret_outs = _replay_aliases_(orig_inputs, fw_outs)
        return ret_outs

RuntimeWrapper orchestration step in isolation (us/call):

| Case | Before (method dispatch) | After (codegen) | Speedup |
|---|---|---|---|
| 0 alias, 0 mut, 5 args | 0.35 us | 0.17 us | 2.1x |
| 2 alias, 0 mut, 5 args | 0.41 us | 0.17 us | 2.5x |
| 0 alias, 2 mut, 5 args | 0.74 us | 0.25 us | 3.0x |
| 3 alias, 1 mut, 10 args | 0.79 us | 0.25 us | 3.2x |
| 5 alias, 3 mut, 20 args | 0.93 us | 0.32 us | 2.9x |

[ghstack-poisoned]