Fix bad merge of triton_heuristics.py

Even though I had picked all upstream changes during merge conflicts,
other parts that didn't have conflicts still picked local changes. Now,
this file is broken with missing symbols. I am just copying the upstream
file into this branch now.

Author: pragupta

torch/_inductor/runtime/triton_heuristics.py

@@ -530,7 +530,7 @@ def _dynamic_scale_rblock(self):
                 #   = regs_per_multiprocessor / (nreg * 32 * num_warps)
                 #   < regs_per_multiprocessor / ((regs_per_multiprocessor / max_threads_per_multi_processor) * 32 * num_warps)
                 #   = max_threads_per_multi_processor / (32 * num_warps)
-                # Using a tigher upper bound can reveal more optimization opportunities.
+                # Using a tighter upper bound can reveal more optimization opportunities.
                 max_blocks_per_sm = max(
                     device_prop.regs_per_multiprocessor // nreg_per_block, 1
                 )
@@ -3241,16 +3241,7 @@ def _persistent_reduction_configs(
         "num_store", 0
     )
 
-    max_autotune_enabled = not disable_pointwise_autotuning(inductor_meta) or (
-        inductor_meta.get("max_autotune")
-        or inductor_meta.get("max_autotune_pointwise")
-    )
-
-    configs = [
-        triton_config_reduction(size_hints, xblock, rnumel, register_intensive=True)
-        for xblock in (1, 8, 32, 128)
-        if xblock == 1 or (xblock <= xnumel and (max_autotune_enabled or rnumel * xblock <= 4096))
-    ]
+    MAX_PERSISTENT_BLOCK_NUMEL = 4096
 
     if triton_meta.get("native_matmul"):
         if len(size_hints) == 3:
@@ -3286,7 +3277,7 @@ def _persistent_reduction_configs(
             )
             for xblock in xblock_vals
             if xblock == 1
-            or (rnumel * xblock <= 4096 and xblock <= xnumel)
+            or (rnumel * xblock <= MAX_PERSISTENT_BLOCK_NUMEL and xblock <= xnumel)
         ]
     else:
         configs = []
@@ -3559,29 +3550,20 @@ def user_autotune(
     )
 
 
-def foreach(triton_meta, filename=None, inductor_meta=None):
+def foreach(triton_meta, num_warps, filename=None, inductor_meta=None):
     """
     Compile a triton foreach kernel
     """
-    configs = []
-    if disable_pointwise_autotuning(inductor_meta) and not (
-        inductor_meta.get("max_autotune") or
-        inductor_meta.get("max_autotune_pointwise")
-    ):
-        configs.append(triton.Config({}, num_stages=1, num_warps=8))
-    else:
-        for warps in [1, 2, 4, 8]:
-            configs.append(triton.Config({}, num_stages=1, num_warps=warps))
-
     return cached_autotune(
         None,
-        configs,
+        [triton.Config({}, num_stages=1, num_warps=num_warps)],
         triton_meta=triton_meta,
         inductor_meta=inductor_meta,
         heuristic_type=HeuristicType.TEMPLATE,
         filename=filename,
     )
 
+
 @dataclasses.dataclass
 class GridExpr:
     """Generate code for grid size expressions in launcher"""