Disable fast math due to precision issue (#10697)

* Disable fast math for fused precision isssue.

* Disable TDU fm.

Author: Zhaowu Pan

slm/model_zoo/gpt-3/external_ops/setup_fp8.py

@@ -58,7 +58,6 @@ def setup_fused_quant_ops():
                     "-DCUTE_ARCH_MMA_SM90A_ENABLE",
                     "--expt-relaxed-constexpr",
                     "--expt-extended-lambda",
-                    "--use_fast_math",
                     "-lineinfo",
                     "-DCUTLASS_DEBUG_TRACE_LEVEL=0",
                     "-maxrregcount=50",
@@ -98,7 +97,6 @@ def setup_token_dispatcher_utils():
                     "-DCUTE_ARCH_MMA_SM90A_ENABLE",
                     "--expt-relaxed-constexpr",
                     "--expt-extended-lambda",
-                    "--use_fast_math",
                     "-maxrregcount=32",
                     "-lineinfo",
                     "-DCUTLASS_DEBUG_TRACE_LEVEL=0",