init llama infer

torchprime/experimental/torchax_models/inference/llama_run.py

@@ -0,0 +1,70 @@
+import torchax.interop
+from torchprime.experimental.torchax_models.inference.llama_run import model
+import torch
+import torchax
+import torchax.config
+import jax
+import time
+
+env = torchax.default_env()
+torch.manual_seed(42)
+torch.set_default_dtype(torch.bfloat16)
+torchax.enable_performance_mode()
+
+max_seq_len = 512  # 8192
+vocab_size = 128  # 32000
+n_layer = 1
+n_heads = 4
+dim = 8
+block_size = 16  # 2048
+batch_size = 1
+
+
+def fake_dataloader(size, vocab_size, seqlen, batch_size):
+  for _ in range(size):
+    x = torch.randint(0, vocab_size, (batch_size, seqlen), device="cpu")
+    yield x
+
+
+if __name__ == "__main__":
+  with torch.no_grad():
+    input = torch.randint(0, vocab_size, (1, max_seq_len))
+    model_args = model.ModelArgs(
+      block_size=block_size,
+      vocab_size=vocab_size,
+      n_layer=n_layer,
+      n_heads=n_heads,
+      dim=dim,
+      max_seq_len=max_seq_len,
+    )
+    freqs_cis = model.precompute_freqs_cis(
+      model_args.dim // model_args.n_heads,
+      model_args.max_seq_len,
+      model_args.rope_theta,
+      model_args.use_scaled_rope,
+    ).to(torch.bfloat16)
+    m = model.Transformer(model_args)
+    m.to(torch.bfloat16)
+
+    # TODO: move weight as arguemts of forward
+    def forward(input, freqs_cis, mask):
+      return m(input, 0, freqs_cis=freqs_cis, mask=mask)
+
+    jitted_forward = torchax.interop.jax_jit(forward)
+
+    data_iter = fake_dataloader(5, vocab_size, max_seq_len, batch_size)
+    with env:
+      m.to("jax")
+      freqs_cis = freqs_cis.to("jax")
+      for i, input in enumerate(data_iter):
+        input = input.to("jax")
+        mask = torch.ones_like(input)
+        step_start = time.perf_counter()
+        output = jitted_forward(input, freqs_cis, mask)
+        jax.block_until_ready(torchax.tensor.t2j(output))
+        step_end = time.perf_counter()
+        print(
+          i,
+          "step latency: ",
+          step_end - step_start,
+        )