[RELEASE-ONLY] Revert buffer ops default enablement on AMD (triton-lang#7456)

Buffer op default enablement is resulting in corresponding PT UT failure
#830

Disabling in release branch for stability while we develop a fix for
main

Author: jataylo

python/test/unit/test_knobs.py

@@ -77,13 +77,10 @@ def test_knobs_scope(fresh_knobs, monkeypatch):
 
     assert fresh_knobs.amd.global_prefetch == 4
     assert fresh_knobs.amd.local_prefetch == 3
-    assert fresh_knobs.amd.use_buffer_ops
 
     # Just to prove that use_buffer_ops is coming from env
     monkeypatch.setenv("AMDGCN_USE_BUFFER_OPS", "0")
     assert not fresh_knobs.amd.use_buffer_ops
-    monkeypatch.delenv("AMDGCN_USE_BUFFER_OPS")
-    assert fresh_knobs.amd.use_buffer_ops
 
     with fresh_knobs.amd.scope():
         fresh_knobs.amd.global_prefetch = 5
@@ -97,13 +94,10 @@ def test_knobs_scope(fresh_knobs, monkeypatch):
 
     assert fresh_knobs.amd.global_prefetch == 4
     assert fresh_knobs.amd.local_prefetch == 3
-    assert fresh_knobs.amd.use_buffer_ops
 
     # Just to prove that use_buffer_ops is coming from env
     monkeypatch.setenv("AMDGCN_USE_BUFFER_OPS", "0")
     assert not fresh_knobs.amd.use_buffer_ops
-    monkeypatch.delenv("AMDGCN_USE_BUFFER_OPS")
-    assert fresh_knobs.amd.use_buffer_ops
 
 
 def test_env_updated(fresh_knobs, monkeypatch):

python/triton/knobs.py

@@ -434,7 +434,7 @@ class nvidia_knobs(base_knobs):
 
 
 class amd_knobs(base_knobs):
-    use_buffer_ops: env_bool = env_bool("AMDGCN_USE_BUFFER_OPS", True)
+    use_buffer_ops: env_bool = env_bool("AMDGCN_USE_BUFFER_OPS")
     dump_amdgcn: env_bool = env_bool("AMDGCN_ENABLE_DUMP")
     libhip_path: env_opt_str = env_opt_str("TRITON_LIBHIP_PATH")
     lld_path: env_opt_str = env_opt_str("TRITON_HIP_LLD_PATH")

third_party/amd/python/test/test_scalarize_packed_fops.py

@@ -64,6 +64,8 @@ def checked_packed_fops_asm_bbs():
 # check scalarization "fixes"
 def test_check_scalarized():
     triton.knobs.amd.scalarize_packed_fops = True
+    triton.knobs.amd.use_buffer_ops = True
+
     kernel = triton.compile(str(Path(__file__).parent / "attn_fwd.ttir"), target=current_target)
 
     # check the specific IR pattern was rewritten