Change datatype to fp16 to run on all GPUs (#13)

* Add comment regarding float16 and bfloat16 data-type compatibility with GPUs

examples/vertex-ai/notebooks/gemma-finetuning-clm-lora-sft.ipynb

@@ -133,7 +133,8 @@
     "    load_in_4bit=True, #  quantize the model to 4-bits when you load it\n",
     "    bnb_4bit_quant_type=\"nf4\", #use a special 4-bit data type for weights initialized from a normal distribution\n",
     "    bnb_4bit_use_double_quant=True, #use a nested quantization scheme to quantize the already quantized weights\n",
-    "    bnb_4bit_compute_dtype=torch.bfloat16, #for faster computation\n",
+    "    bnb_4bit_compute_dtype=torch.bfloat16, #conversion from bfloat16 to float16 may lead to overflow (and opposite may lead to loss of precision)\n",
+    "                                           # Use float16 when running on a GPU(T4, V100) where bfloat16 is not supported\n",
     ")"
    ]
   },
@@ -178,7 +179,7 @@
    "source": [
     "## Load the tokenizer\n",
     "# tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
-    "tokenizer = GemmaTokenizer.from_pretrained(model_id.replace(\"gemma\", \"golden-gate\")) # Until the tokenizer is available in the model hub"
+    "tokenizer = GemmaTokenizer.from_pretrained(model_id)"
    ]
   },
   {
@@ -198,7 +199,7 @@
    },
    "outputs": [],
    "source": [
-    "from peft import get_peft_model, LoraConfig, TaskType\n",
+    "from peft import LoraConfig\n",
     "\n",
     "peft_config = LoraConfig(\n",
     "    task_type=\"Causal_LM\", \n",
@@ -264,7 +265,8 @@
     "    group_by_length=True,\n",
     "    eval_steps=20, # Evaluate every n steps during training\n",
     "    evaluation_strategy=\"steps\",\n",
-    "    bf16=True\n",
+    "    bf16=True  #conversion from bfloat16 to float16 may lead to overflow (and opposite may lead to loss of precision)\n",
+    "               # Use float16 when running on a GPU(T4, V100) where bfloat16 is not supported\n",
     ")\n",
     "\n",
     "trainer = SFTTrainer(\n",