Update app.py

app.py

@@ -5,11 +5,13 @@ import gradio as gr
 
 from src.client import DistributedBloomForCausalLM
 
-INITIAL_PEERS = [
-'/ip4/193.106.95.184/tcp/31337/p2p/QmUigSxrVz9x5FR9ZYr4iRfEX2vDxihL2YZtDd7sp2eKnM',
-'/ip6/193.106.95.184/tcp/21337/p2p/QmSXDXLeSMXjS4YerDrdn1zpGQaNzkZ9ogN2SoAEyAdDhs',
-'/ip6/193.106.95.184/udp/21337/quic/QmSXDXLeSMXjS4YerDrdn1zpGQaNzkZ9ogN2SoAEyAdDhs',
-]
+INITIAL_PEERS = ['/ip4/193.106.95.184/tcp/443/p2p/QmSXDXLeSMXjS4YerDrdn1zpGQaNzkZ9ogN2SoAEyAdDhs']
+
+import hivemind  # test that DHT instances work on localhost
+dht1 = hivemind.DHT(start=True)
+dht2 = hivemind.DHT(start=True, initial_peers=dht1.get_visible_maddrs())
+
+
 tokenizer = transformers.BloomTokenizerFast.from_pretrained("bigscience/test-bloomd-6b3")
 model = DistributedBloomForCausalLM.from_pretrained("bigscience/test-bloomd-6b3", initial_peers=INITIAL_PEERS, low_cpu_mem_usage=True, torch_dtype=torch.float32)
 
@@ -19,16 +21,8 @@ def inference(text, seq_length=1):
         for i in range(seq_length):
             h = model.transformer.word_embeddings(input_ids)
             h = model.transformer.word_embeddings_layernorm(h)
-
-            h = remote_transformer.step(h)  # note [yozh]: this line currently freezes for 10 seconds first time only, its gonna be fixed in the nearest PR
-
-            h = model.transformer.ln_f(h)
-            h = F.linear(h, weight=model.transformer.word_embeddings.weight)  # note: this line takes a while, will also be fixed
-            next_token_ix = torch.multinomial((h[0, -1] / 0.8).softmax(-1), 1)
-            
-            # print(end=tokenizer.decode(next_token_ix.item()))
-            input_ids = next_token_ix.view(1, 1)
-    return tokenizer.decode(input_ids.item())
-
+            h = remote_transformer.step(h)
+            return repr(h)
+    
 iface = gr.Interface(fn=inference, inputs="text", outputs="text")
 iface.launch()