Update README.md

README.md

@@ -1,3 +1,67 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+---
+
+Sample Code
+```
+import requests
+import torch
+import os
+from tqdm import tqdm
+from transformers import AutoProcessor, Gemma3ForConditionalGeneration, AutoTokenizer
+
+
+model_id = "ai4bharat/IndicTrans3-gemma-beta"
+language = "Hindi"
+# permitted languages = Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Sanskrit, Tamil, Telugu, Urdu
+
+model = Gemma3ForConditionalGeneration.from_pretrained(
+    model_id, device_map="auto", attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16
+).eval()
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+src = [
+    "When I was young, I used to go to the park every day.",
+    "We watched a new movie last week, which was very inspiring.",
+    "If you had met me at that time, we would have gone out to eat.",
+    "My friend has invited me to his birthday party, and I will give him a gift."
+]
+
+_PROMPT = (
+    "<bos><start_of_turn>user\n"
+    "Translate the following text to {tgt_lang}: {source_text}:"
+    "<end_of_turn>\n<start_of_turn>model\n"
+)
+
+batch_size = 100  # Adjust based on memory constraints
+outputs = []
+
+for i in tqdm(range(0, len(src), batch_size)):
+    batch = src[i:i + batch_size]
+
+    batch = [
+        _PROMPT.format(
+            tgt_lang=language,
+            source_text=s
+        )
+        for s in batch
+    ]
+    tokinp = tokenizer(batch, return_tensors='pt', padding="longest")
+    for k in tokinp:
+        tokinp[k] = tokinp[k].to("cuda")
+
+    out = model.generate(
+        **tokinp,
+        max_new_tokens=8192,
+        num_beams=1,
+        do_sample=False
+    )
+
+    for b, o in zip(batch, out):
+        input_length = len(tokenizer(b)['input_ids'])
+        finout = tokenizer.decode(o, skip_special_tokens=True)
+        outputs.append(finout.split('model')[-1].strip())
+
+print(outputs)
+```