bump pytorch dep to and docker image to 2.8.0-rc3

Author: speediedan

tests/test_model_parallel.py

@@ -516,10 +516,11 @@ def __post_init__(self):
     ModParallelTestCfg(model_cfg_key="tp_lp_fp16", model_cls=cm_mod_parallel, precision_opts=fp16, model_cfg=tp_only,
                        strategy_cfg=dp1_tp2, runif_alias="alone",
                        expected_results=ExpectedResults(expected_state=path_tp)),
-    # TODO: temporarly disabling until issue addressed in upstream RC
+    # TODO: given that SDPBackend.MATH with lp causes (at least without special accommodation)
+    #       https://github.com/pytorch/pytorch/pull/149764 as of PT 2.8, we are disabling this test until our
+    #       infrastructure is upgraded and we can use SDPBackend.FLASH with bf16 on all our cards
     # ModParallelTestCfg(model_cfg_key="tp_lp_bf16", model_cls=cm_mod_parallel, precision_opts=bf16,
     #                    model_cfg=tp_lp_math_sdp_impl, strategy_cfg=dp1_tp2, runif_alias="bf16_alone"),
-
     # FSDP2 + TP (trivial submesh) tests
     # temporarily disabling this test as it triggers a hang in `fsdp_autocm_tp` about 10% of the time and
     # `fsdp_autocm_tp` and the marginal utility of `fsdp_tp` is minimal