Drop AutoWS warp-spec kwargs from Triton autotune configs (#11)

Summary:
Pull Request resolved: #11

Remove dead, unused Triton autotune configs: the manual/warp-specialization tuning kwargs that are no longer needed —
- num_buffers_warp_spec
- num_consumer_groups
- reg_dec_producer
- reg_inc_consumer

These lowercase kwargs were passed to triton.Config only inside the warp-spec (WS) autotune lists, which are selected solely when enable_ws=True. Production never enables WS (the single fronting module efficient_module_suite/modules/generalized_dot_product_attention.py -> GeneralizedDotProductAttention calls every Triton path with enable_ws=False), so these configs are dead/unused in prod. reg_dec_producer/reg_inc_consumer do not appear in these files.

Confirmed the retained paths ARE necessary and were left untouched: the uppercase NUM_CONSUMER_GROUPS constexpr in the config meta dicts is a real kernel parameter — it drives tiling math (BLOCK_M // NUM_CONSUMER_GROUPS) and, in _dense, the kernels' manual warp specialization (tl.async_task([1, NUM_CONSUMER_GROUPS]), 19 sites). That manual-WS path is independent of the removed autotune knobs, so only the dead lowercase Config kwargs are dropped (plus now-dead 'for buf in [2]' clauses that only fed num_buffers_warp_spec).

Files (7): triton_generalized_dot_product_attention.py, _dense, triton_sdyt_gdpa.py, triton_sdyt_gdpa_add.py, triton_matmul_layernorm_persistent.py, triton_jfa_v3.py, ops/oss/gdpa/triton/triton_generalized_dot_product_attention.py. No changes in amd/mtia variants (uppercase constexpr only) or triton_jagged_dense_bmm_gelu_rms.py (commented out).

Reviewed By: santoshmo

Differential Revision: D111169324

fbshipit-source-id: 42f00df2ba74ea1f918238d862a2f734736ea3cf

Author: njriasan

gdpa/triton/triton_generalized_dot_product_attention.py

@@ -358,8 +358,6 @@ def _gdpa_fwd_inner_ws(
             {"BLOCK_M": BM, "BLOCK_N": BN, "NUM_CONSUMER_GROUPS": 2},
             num_stages=s,
             num_warps=w,
-            num_buffers_warp_spec=2,
-            num_consumer_groups=2,
         )
     )
     # pyrefly: ignore [not-iterable]
@@ -1846,11 +1844,8 @@ def _gdpa_bwd_dq(
             },
             num_stages=s,
             num_warps=w,
-            num_buffers_warp_spec=buf,
-            num_consumer_groups=2,
         )
     )
-    for buf in [2]
     # pyrefly: ignore [not-iterable]
     for BM1 in block_m_hw_supported([64])
     # pyrefly: ignore [not-iterable]