fix an error in gemma accuracy example (#351)

Author: jianyizh

example/gemma/README.md

@@ -131,26 +131,28 @@ Based on lm-eval-harness.
 | **Parameter** | **Default Value** |
 | :---: | :--- |
 | **model** | gemma |
-| **model_name** | gemma_2b_en |
-| **dtype** | bfloat16 |
-| **num_beams** | 1 |
+| **model_name** | **gemma_2b_en**,  gemma_7b_en| 
+| **dtype** | **bfloat16**, float16, float32 |
+| **num_beams** | **1** |
 ```
 git clone https://github.com/EleutherAI/lm-evaluation-harness.git lm_eval
 cd lm_eval
 git checkout b281b0921b636bc36ad05c0b0b0763bd6dd43463
 git apply ../gemma.patch
 pip install -r requirements.txt
+pip install torch --index-url https://download.pytorch.org/whl/cpu --force-reinstall
+export KERAS_BACKEND=jax
 python main.py \
   --model gemma \
-  --model_args model_name=gemma_2b_en,dtype=float32,num_beams=1 \
+  --model_args model_name=gemma_7b_en,dtype=bfloat16,num_beams=4 \
   --tasks openbookqa \
   --no_cache
 ```
 ### Output
 ```
-gemma (model_name=gemma_2b_en,dtype=float32,num_beams=1), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
+gemma (model_name=gemma_7b_en,dtype=bfloat16,num_beams=4), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
 |   Task   |Version| Metric |Value|   |Stderr|
 |----------|------:|--------|----:|---|-----:|
-|openbookqa|      0|acc     |0.302|±  |0.0206|
-|          |       |acc_norm|0.398|±  |0.0219|
+|openbookqa|      0|acc     |0.326|±  |0.0210|
+|          |       |acc_norm|0.454|±  |0.0223|
 ```

example/gemma/gemma.patch

@@ -20,7 +20,7 @@ index 8ca27fac..6127ce6c 100644
 
 diff --git a/lm_eval/models/gemma.py b/lm_eval/models/gemma.py
 new file mode 100644
-index 00000000..bc4540f7
+index 00000000..732185c4
 --- /dev/null
 +++ b/lm_eval/models/gemma.py
 @@ -0,0 +1,79 @@
@@ -41,7 +41,7 @@ index 00000000..bc4540f7
 +        self.model = keras_nlp.models.GemmaCausalLM.from_preset(model_name)
 +        if num_beams > 1:
 +            from keras_nlp.samplers import BeamSampler
-+            model.compile(sampler=BeamSampler(num_beams=args.num_beams))
++            self.model.compile(sampler=BeamSampler(num_beams=num_beams))
 +
 +    @property
 +    def eot_token_id(self):