diff --git a/.github/workflows/amd.yml b/.github/workflows/amd.yml
index 97b96f2d5875..975339952c12 100644
--- a/.github/workflows/amd.yml
+++ b/.github/workflows/amd.yml
@@ -28,6 +28,7 @@ jobs:
       # Runs a single command using the runners shell
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           rocm-smi --showhw
           which python
           python --version
diff --git a/.github/workflows/nv-accelerate-v100.yml b/.github/workflows/nv-accelerate-v100.yml
index cc1abfc5717d..436e11b3f822 100644
--- a/.github/workflows/nv-accelerate-v100.yml
+++ b/.github/workflows/nv-accelerate-v100.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-inference.yml b/.github/workflows/nv-inference.yml
index 681fc90be78f..dc00682edd46 100644
--- a/.github/workflows/nv-inference.yml
+++ b/.github/workflows/nv-inference.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
@@ -40,7 +41,7 @@ jobs:
           git clone https://github.com/huggingface/transformers
           cd transformers
           # if needed switch to the last known good SHA until transformers@master is fixed
-          # git checkout 1cc453d33
+          git checkout v4.21.2
           git rev-parse --short HEAD
           pip uninstall --yes transformers
           pip install .
@@ -61,4 +62,5 @@ jobs:
           if [[ -d ./torch-extensions ]]; then rm -rf ./torch-extensions; fi
           cd tests
           EXPECTED_TORCH=$(pip index versions torch | grep -oP -m1 "^\s*LATEST.*\s\K\d+\.\d+")
+          TRANSFORMERS_CACHE=/blob/transformers_cache/ TORCH_EXTENSIONS_DIR=./torch-extensions pytest --color=yes --durations=0 --verbose -m 'seq_inference' unit/ --torch_ver=$EXPECTED_TORCH --cuda_ver="11.3"
           TRANSFORMERS_CACHE=/blob/transformers_cache/ TORCH_EXTENSIONS_DIR=./torch-extensions pytest --color=yes --durations=0 -n 4 --verbose -m 'inference' unit/ --torch_ver=$EXPECTED_TORCH --cuda_ver="11.3"
diff --git a/.github/workflows/nv-lightning-v100.yml b/.github/workflows/nv-lightning-v100.yml
index 9fe1043b96d0..a63f6b75e769 100644
--- a/.github/workflows/nv-lightning-v100.yml
+++ b/.github/workflows/nv-lightning-v100.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-nightly.yml b/.github/workflows/nv-nightly.yml
index 98afd75105a3..40abe198c7cd 100644
--- a/.github/workflows/nv-nightly.yml
+++ b/.github/workflows/nv-nightly.yml
@@ -17,6 +17,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-torch-latest-v100.yml b/.github/workflows/nv-torch-latest-v100.yml
index 1dc535f9b327..c8a3c32bdc36 100644
--- a/.github/workflows/nv-torch-latest-v100.yml
+++ b/.github/workflows/nv-torch-latest-v100.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-torch-nightly-v100.yml b/.github/workflows/nv-torch-nightly-v100.yml
index e1c916afba2d..d456483c72f5 100644
--- a/.github/workflows/nv-torch-nightly-v100.yml
+++ b/.github/workflows/nv-torch-nightly-v100.yml
@@ -17,6 +17,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-torch12-p40.yml b/.github/workflows/nv-torch12-p40.yml
index 944ba3beb19d..29b9f891c3bb 100644
--- a/.github/workflows/nv-torch12-p40.yml
+++ b/.github/workflows/nv-torch12-p40.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-torch18-v100.yml b/.github/workflows/nv-torch18-v100.yml
index b512ea29113f..2971416ebd81 100644
--- a/.github/workflows/nv-torch18-v100.yml
+++ b/.github/workflows/nv-torch18-v100.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/.github/workflows/nv-transformers-v100.yml b/.github/workflows/nv-transformers-v100.yml
index 945457b304c5..bfc9919be1f9 100644
--- a/.github/workflows/nv-transformers-v100.yml
+++ b/.github/workflows/nv-transformers-v100.yml
@@ -24,6 +24,7 @@ jobs:
 
       - name: environment
         run: |
+          echo "JobID: $AISC_NODE_INSTANCE_ID"
           nvidia-smi
           which python
           python --version
diff --git a/tests/pytest.ini b/tests/pytest.ini
index a52a49e5bbc3..b7ee315be801 100644
--- a/tests/pytest.ini
+++ b/tests/pytest.ini
@@ -1,6 +1,7 @@
 [pytest]
-addopts = -m "not sequential and not nightly and not inference"
+addopts = -m "not sequential and not nightly and not inference and not seq_inference"
 markers =
     sequential:Tests that need to be run sequentially
     inference:Inference model tests
+    seq_inference:Inference model tests to run sequentially
     nightly:Tests that should be run nightly
diff --git a/tests/unit/inference/test_inference.py b/tests/unit/inference/test_inference.py
index cdc3e83232de..1b1efdc595fe 100644
--- a/tests/unit/inference/test_inference.py
+++ b/tests/unit/inference/test_inference.py
@@ -52,7 +52,7 @@ def lm_eval_imports():
     "distilgpt2",
     "Norod78/hebrew-bad_wiki-gpt_neo-tiny",
     "EleutherAI/gpt-j-6B",
-    "bigscience/bloom-350m",
+    "bigscience/bloom-560m",
 ]
 _opt_models = [
     "facebook/opt-125m",        # 125m, 1.7B, ..., 175B variants have the same model architecture.
@@ -111,6 +111,7 @@ def enable_cuda_graph(request):
 @pytest.fixture()
 def invalid_model_task_config(model_w_task, dtype, enable_cuda_graph):
     model, task = model_w_task
+    msg = ""
     if pkg_version.parse(torch.__version__) <= pkg_version.parse("1.2"):
         msg = "DS inference injection doesn't work well on older torch versions"
     elif model not in pytest.all_models[task]:
@@ -120,10 +121,17 @@ def invalid_model_task_config(model_w_task, dtype, enable_cuda_graph):
     elif enable_cuda_graph and pkg_version.parse(
             torch.__version__) < pkg_version.parse("1.10"):
         msg = "CUDA Graph is only available in torch versions >= 1.10"
-    elif ("gpt-j-6B" in model) and (dtype == torch.float):
+    elif "gpt-j-6B" in model:
+        if dtype != torch.half:
+            msg = f"Not enough GPU memory to run {model} with dtype {dtype}"
+        elif enable_cuda_graph:
+            msg = f"Not enough GPU memory to run {model} with CUDA Graph enabled"
+    elif "gpt-neox-20b" in model:  # TODO: remove this when neox issues resolved
+        msg = "Skipping gpt-neox-20b for now"
+    elif ("gpt-neox-20b" in model) and (dtype != torch.half):
         msg = f"Not enough GPU memory to run {model} with dtype {dtype}"
-    else:
-        msg = ""
+    elif ("bloom" in model) and (dtype != torch.half):
+        msg = f"Bloom models only support half precision, cannot use dtype {dtype}"
     return msg
 
 
@@ -160,7 +168,7 @@ def query(model_w_task):
 def inf_kwargs(model_w_task):
     model, task = model_w_task
     if task == "text-generation":
-        return {"do_sample": False}
+        return {"do_sample": False, "max_length": 20}
     else:
         return {}
 
@@ -228,7 +236,9 @@ def test(
         local_rank = int(os.getenv("LOCAL_RANK", "0"))
 
         if "gpt-j-6B" in model and dtype == torch.half:
-            _model = AutoModelForCausalLM.from_pretrained(model)
+            _model = AutoModelForCausalLM.from_pretrained(model,
+                                                          revision="float16",
+                                                          torch_dtype=torch.float16)
             tokenizer = AutoTokenizer.from_pretrained(model)
             _model.half()
             pipe = pipeline(
@@ -269,7 +279,9 @@ def test(
         torch.cuda.synchronize()
         ds_time = time.time() - start
 
-        if task == "text-generation":
+        # facebook/opt* and some bigscient/bloom* models are not matching
+        # baseline exactly, adding an exception to them for now
+        if ("opt" in model) or ("bloom" in model):
             bs_output = pipe(query, **inf_kwargs)
 
         # These performance tests are only measuring the time for a single
@@ -278,6 +290,58 @@ def test(
         assert assert_fn(bs_output, ds_output)
 
 
+@pytest.mark.seq_inference
+@pytest.mark.parametrize("model_w_task",
+                         [("gpt2",
+                           "text-generation"),
+                          ("EleutherAI/gpt-neox-20b",
+                           "text-generation"),
+                          ("bigscience/bloom-3b",
+                           "text-generation")],
+                         ids=["gpt2",
+                              "gpt-neox",
+                              "bloom"])
+class TestMPSize(DistributedTest):
+    world_size = 4
+
+    def test(
+        self,
+        model_w_task,
+        dtype,
+        enable_cuda_graph,
+        query,
+        inf_kwargs,
+        assert_fn,
+        invalid_model_task_config,
+    ):
+        if invalid_model_task_config:
+            pytest.skip(invalid_model_task_config)
+
+        model, task = model_w_task
+        local_rank = int(os.getenv("LOCAL_RANK", "0"))
+
+        # We have to load these large models on CPU with pipeline because not
+        # enough GPU memory
+        pipe = pipeline(task, model=model, device=-1, framework="pt")
+        bs_output = pipe(query, **inf_kwargs)
+
+        pipe.model = deepspeed.init_inference(
+            pipe.model,
+            mp_size=self.world_size,
+            dtype=dtype,
+            replace_method="auto",
+            replace_with_kernel_inject=True,
+            enable_cuda_graph=enable_cuda_graph,
+        )
+        # Switch device to GPU so that input tensors are not on CPU
+        pipe.device = torch.device(f"cuda:{local_rank}")
+        ds_output = pipe(query, **inf_kwargs)
+
+        print(local_rank, "baseline", bs_output)
+        print(local_rank, "deepspeed", ds_output)
+        assert assert_fn(bs_output, ds_output)
+
+
 @pytest.mark.nightly
 @pytest.mark.parametrize(
     "model_family, model_name",