FIX Use trl version of tiny random llama (#1681)

Using the version from HuggingFaceM4 broke our tests because it was
updated. Although the update is reverted, we still better switch to this
version, which is explicitly for testing and should be stable.

tests/test_common_gpu.py

@@ -469,7 +469,7 @@ def test_lora_seq2seq_lm_multi_gpu_inference(self):
     @require_bitsandbytes
     def test_adaption_prompt_8bit(self):
         model = LlamaForCausalLM.from_pretrained(
-            "HuggingFaceM4/tiny-random-LlamaForCausalLM",
+            "trl-internal-testing/tiny-random-LlamaForCausalLM",
             quantization_config=BitsAndBytesConfig(load_in_8bit=True),
             torch_dtype=torch.float16,
             device_map="auto",
@@ -492,7 +492,7 @@ def test_adaption_prompt_8bit(self):
     @require_bitsandbytes
     def test_adaption_prompt_4bit(self):
         model = LlamaForCausalLM.from_pretrained(
-            "HuggingFaceM4/tiny-random-LlamaForCausalLM",
+            "trl-internal-testing/tiny-random-LlamaForCausalLM",
             quantization_config=BitsAndBytesConfig(load_in_4bit=True),
             torch_dtype=torch.float16,
             device_map="auto",
@@ -982,7 +982,7 @@ def test_4bit_dora_merging(self):
             bnb_4bit_compute_dtype=torch.float32,
         )
         model = AutoModelForCausalLM.from_pretrained(
-            "HuggingFaceM4/tiny-random-LlamaForCausalLM",
+            "trl-internal-testing/tiny-random-LlamaForCausalLM",
             quantization_config=bnb_config,
             torch_dtype=torch.float32,
         ).eval()

tests/test_decoder_models.py

@@ -32,7 +32,7 @@
     "hf-internal-testing/tiny-random-gpt_neo",
     "hf-internal-testing/tiny-random-GPTJForCausalLM",
     "hf-internal-testing/tiny-random-GPTBigCodeForCausalLM",
-    "HuggingFaceM4/tiny-random-LlamaForCausalLM",
+    "trl-internal-testing/tiny-random-LlamaForCausalLM",
 ]
 
 FULL_GRID = {
@@ -340,7 +340,7 @@ def test_passing_input_embeds_works(self, test_name, model_id, config_cls, confi
         self._test_passing_input_embeds_works(test_name, model_id, config_cls, config_kwargs)
 
     def test_lora_layer_replication(self):
-        model_id = "HuggingFaceM4/tiny-random-LlamaForCausalLM"
+        model_id = "trl-internal-testing/tiny-random-LlamaForCausalLM"
         config_kwargs = {
             "target_modules": ["down_proj", "up_proj"],
             "task_type": "CAUSAL_LM",