Align reduce precision with PyTorch 2.9.1, disable precision checks.

[baoqiwen]

PR types

Others

PR changes

Others

Description

reduce 精度对齐 torch 2.9.1，影响一系列 llama & llm case，注释掉这些 case 的精度测试。

V100

	loss_base	loss_test	diff
llama_dygraph_auto_bs8_fp32_DP2	9.4992733	9.49927235	9.5E-07
llama_dygraph_auto_bs8_fp32_DP2-MP2	9.3507843	9.35078526	-9.6E-07
llama_dygraph_auto_bs8_fp16_DP2-MP2-PP2	9.35162163	9.35163498	-1.335E-05
llama_align_dygraph_dy2st_auto_bs2_bf16_DP2-MP1-PP1	9.99302597	9.99302673	-7.6E-07
llama_pir_auto_fuse_ffn_attention_qkv_MP2	9.49613285	9.4961319	9.5E-07
llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP2-SP	9.25199432	9.2519928	1.52E-06
llama_align_dy2st_fthenb_and_vpp_auto_bs2_fp32_DP1-MP1-PP4	10.24240494	10.24240398	9.6E-07
llama_dy2st_auto_bs2_bf16_DP2-MP1-PP1-CINN	9.99302597	9.99302673	-7.6E-07
llm_gpt_dygraph_auto_bs8_fp32_DP2	10.55727577	10.55727673	-9.6E-07
llm_gpt_recompute_bs32_bf16_MP2-SD4-stage1	8.93362617	8.93362999	-3.82E-06

A100

	loss_base	loss_test	diff
llama_dygraph_auto_bs4_bf16_SD2	9.23504105	9.23507309	-3.204E-05
llama_dygraph_auto_bs8_fp32_DP2-MP2	9.38577747	9.38577652	9.5E-07
llama_dygraph_auto_bs8_fp16_DP2-MP2-PP2	9.39368343	9.39368439	-9.6E-07
llama_dygraph_auto_bs8_fp16_DP2-MP2-CP2	9.38431168	9.38274002	0.00157166
llama_dygraph_auto_bs8_fp16_DP2-MP2-PP2_hybrid_pp	9.57190609	9.57185268	5.341E-05
llama_align_dygraph_dy2st_auto_bs2_bf16_DP2-MP1-PP1	10.20990601	10.20989227	1.374E-05
llama_pir_auto_fuse_ffn_attention_qkv_MP2	10.58283806	10.58283901	-9.5E-07
llama_convert_hybrid_ckpt_to_auto_parallel_bs2_fp32_DP2-MP1-PP1	11.004673	11.00467396	-9.6E-07
llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP1-SP	9.37980728	9.37988205	-7.477E-05
llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP2-SP	9.44232788	9.44226913	5.875E-05
llama_dy2st_auto_bs2_bf16_DP2-MP1-PP1-CINN	10.20990143	10.20990906	-7.63E-06
llama_lora_static_graph_auto_bs_2_bf16_DP2-TP2-PP1	14.08647537	14.08639145	8.392E-05

[paddle-bot]

Thanks for your contribution!

[baoqiwen]

/re-run all-failed