diff --git a/nlp/gpt_j/popxl/README.md b/nlp/gpt_j/popxl/README.md
index 4a6b230ca..2cec6ed06 100644
--- a/nlp/gpt_j/popxl/README.md
+++ b/nlp/gpt_j/popxl/README.md
@@ -1,10 +1,9 @@
 # GPT-J
 GPT-J for NLP pre-training and text generation, optimised for Graphcore's IPU.
 
-| Framework | domain | Model | Datasets | Tasks| Training| Inference |
-|-------------|-|------|-------|-------|-------|---|
-| opXL | NLP | GPT-J | MNLI | Next sentence prediction, Question/Answering | ✅  | ✅ |
-
+| Framework | Domain | Model | Datasets | Tasks | Training | Inference |
+|-----------|--------|-------|----------|-------|----------|-----------|
+| popXL | NLP | GPT-J | MNLI | Next sentence prediction, Question/Answering | | <p style="text-align: center;">✅ <br> Min. 16 IPUs (POD16) required | <p style="text-align: center;">✅ <br> Min. 16 IPU (POD16) required |
 
 # Instructions summary
 
@@ -51,7 +50,6 @@ source <venv path>/bin/activate
 3. Install the Python requirements:
 ```bash
 pip3 install -r requirements.txt
-```nstall
 ```
 
 
@@ -72,9 +70,10 @@ The task is to predict the relation between the premise and the hypothesis, whic
 
 
 The default model size for fine-tuning is GPT-J 6B on POD64 (named `gptj_6B_1024_pod64`). You can
-change it to other configurations that are available in the configuration file `config/finetuning.yml` using the `- -config` CLI parameter:
+change it to other configurations that are available in the configuration file `config/finetuning.yml` using the `--config` CLI parameter.
+In particular, you can run fine-tuning on a POD16 using
 ```bash
-python3 run_finetuning_mnli.py - -config gptj_6B_1024_pod64
+python3 run_finetuning_mnli.py --config gptj_6B_1024_pod16
 ```
 
 When running the application, it is possible to save/load executables to/from a cache store. This allows for reusing a saved executable instead of re-compiling the model when re-running identical model configurations. To enable this, use the environment variable `POPXL_CACHE_DIR=<PATH/TO/CACHE>` when running the application:
diff --git a/nlp/gpt_j/popxl/config/finetuning_mnli.yml b/nlp/gpt_j/popxl/config/finetuning_mnli.yml
index b2b97f879..3d3afb594 100644
--- a/nlp/gpt_j/popxl/config/finetuning_mnli.yml
+++ b/nlp/gpt_j/popxl/config/finetuning_mnli.yml
@@ -57,6 +57,17 @@ release:
       available_memory_proportion: [ 0.2 ]
       attention_serialisation: 2
 
+  "gptj_6B_1024_pod16":
+    <<: *gptj_6B_1024
+    execution:
+      micro_batch_size: 1
+      loss_scaling: 4096
+      io_tiles: 128
+      data_parallel: 1
+      tensor_parallel: 16
+      available_memory_proportion: [ 0.2 ]
+      attention_serialisation: 2
+
   tiny:
     <<: *tiny
     execution:
diff --git a/nlp/gpt_j/popxl/inference.py b/nlp/gpt_j/popxl/inference.py
index 4366dd325..86d41f342 100644
--- a/nlp/gpt_j/popxl/inference.py
+++ b/nlp/gpt_j/popxl/inference.py
@@ -70,10 +70,10 @@ def inference(config: GPTJConfig) -> TaskSession:
             # ----- Create Variables -----
 
             # Create RemoteBuffers for each variable
-            embeddings_buffers = named_variable_buffers(embeddings_facts)
+            embeddings_buffers = named_variable_buffers(embeddings_facts, shard_over_dict=False)
             layer_buffers = named_variable_buffers(
-                layer_facts, entries=config.model.layers)
-            lm_buffers = named_variable_buffers(lm_facts)
+                layer_facts, entries=config.model.layers, shard_over_dict=False)
+            lm_buffers = named_variable_buffers(lm_facts, shard_over_dict=False)
 
             variables = NamedTensors()
             transformer = NamedTensors()
diff --git a/nlp/gpt_j/popxl/requirements.txt b/nlp/gpt_j/popxl/requirements.txt
index 5b594dd4f..cafd2b135 100644
--- a/nlp/gpt_j/popxl/requirements.txt
+++ b/nlp/gpt_j/popxl/requirements.txt
@@ -15,6 +15,6 @@ sklearn==0.0
 pytest==6.2.5
 pytest-pythonpath==0.7.4
 
-git+ssh://git@github.com/graphcore/popxl-addons.git@sdk-release-3.1
+git+ssh://git@github.com/graphcore/popxl-addons.git@sdk-release-3.1_a
 
 protobuf==3.20.*; python_version > '3.6'