v1

app.py

@@ -1,7 +1,30 @@
+import torch
+import torch.nn.functional as F
+import transformers
 import gradio as gr
 
-def greet(name):
-    return "Hello " + name + "!!"
+from src.client import DistributedBloomForCausalLM
 
-iface = gr.Interface(fn=greet, inputs="text", outputs="text")
+INITIAL_PEERS = ['/ip4/193.106.95.184/tcp/31337/p2p/QmUigSxrVz9x5FR9ZYr4iRfEX2vDxihL2YZtDd7sp2eKnM']
+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)
+
+def inference(text, seq_length=1):
+    input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+    with torch.inference_mode(), model.transformer.h.inference_session() as remote_transformer:
+        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())
+
+iface = gr.Interface(fn=inference, inputs="text", outputs="text")
 iface.launch()