clean up temp, experimental codes

Signed-off-by: cliu-us <[email protected]>

Author: chichun-charlie-liu

examples/QAT_INT8/run_qa_no_trainer_qat.py

@@ -393,12 +393,6 @@ def parse_args():
         default=None,
         help="convert QAT model to utilize real INT8 GPU kernel, 'cutlass' or 'triton'",
     )
-    parser.add_argument(
-        "--verify_preccfg",
-        type=str,
-        default=None,
-        help="Path to an old version of AIU checkpoint, which should have a precconfig.json",
-    )
 
     args = parser.parse_args()
 
@@ -1201,18 +1195,6 @@ def speedtest(model, exam_inp, Ntest=100):
 
         return
 
-    elif args.verify_preccfg:
-        # Local
-        from fms_mo.utils.aiu_utils import verify_preccfg
-
-        exam_inp = next(iter(train_dataloader))
-
-        model_aiu = verify_preccfg(args.verify_preccfg, exam_inp, sim_level=3)
-
-        metrics = squad_eval(model_aiu)
-        logger.info(metrics)
-        return
-
     # ----------------------------------------------------
 
     # Only show the progress bar once on each machine.

fms_mo/custom_ext_kernels/triton_kernels.py

@@ -235,8 +235,7 @@ def imatmul_kernel(
     accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.int32)
     ## ------ prepare LSB rounding/truncation masks -------
     round_bit = 1 << (chunk_trun_bits - 1) if chunk_trun_bits > 0 else 0
-    # full_32b_mask = 0xFFFFFFFF
-    # trun_mask = (full_32b_mask << chunk_trun_bits) & full_32b_mask
+    msb_mask = 0x00FFFFFF  # only needed when simulating truncation on MSB
     ## ---------------------------------------------------------
 
     for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):