Update gemma2 examples with a note about sample generation (#1176)

SUMMARY:
- Add a note advising users to either downgrade transformers from 4.49
or use vLLM for generation
- We should revisit why this is only happening on generation with this
new release but can be revisited down the road

Author: dsikka

examples/quantization_kv_cache/gemma2_fp8_kv_example.py

@@ -86,6 +86,10 @@ def process_and_tokenize(example):
     "Please use vLLM for inference with the quantized kv_cache.",
 )
 # Confirm generations of the quantized model look sane.
+
+# NOTE: transformers 4.49.0 results in a generation error with gemma2.
+# Consider either downgrading your transformers version to a previous version
+# or use vLLM for sample generation.
 print("\n\n")
 print("========== SAMPLE GENERATION ==============")
 input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")

examples/quantization_w8a8_fp8/gemma2_example.py

@@ -29,6 +29,9 @@
 )
 
 # Confirm generations of the quantized model look sane.
+# NOTE: transformers 4.49.0 results in a generation error with gemma2.
+# Consider either downgrading your transformers version to a previous version
+# or use vLLM for sample generation.
 print("========== SAMPLE GENERATION ==============")
 input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
 output = model.generate(input_ids, max_new_tokens=20)

examples/quantization_w8a8_int8/gemma2_example.py

@@ -68,6 +68,9 @@ def tokenize(sample):
 )
 
 # Confirm generations of the quantized model look sane.
+# NOTE: transformers 4.49.0 results in a generation error with gemma2.
+# Consider either downgrading your transformers version to a previous version
+# or use vLLM for sample generation.
 print("========== SAMPLE GENERATION ==============")
 input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
 output = model.generate(input_ids, max_new_tokens=20)