Merge pull request #1289 from AI-Hypercomputer:sixiang/debug

PiperOrigin-RevId: 730570138

MaxText/inference_mlperf/offline_inference.py

@@ -340,10 +340,10 @@ def decode():
         else:
           assert False, "no generate fn"
         result_tokens_l = []
-        for i in range(5):
+        for i in range(10):
           self.decode_state, result_tokens = gen_fn(self.params, self.decode_state, None)
           result_tokens_l.append(result_tokens)
-      for i in range(5):
+      for i in range(10):
         # result_tokens.copy_to_host_async()
         result_tokens = result_tokens_l[i].convert_to_numpy()
         self.detokenize_backlog.put((result_tokens, False, 0, 0), block=True)
@@ -414,7 +414,7 @@ def detokenize():
         self.detokenize_backlog.put((first_token, True, row.id, slot), block=True)
         continue
 
-      if len(self.prefill_buckets[padded_len // 2]) == 0:
+      if len(self.prefill_buckets[padded_len // 2]) != 0:
         prefill_batch(self.prefill_buckets[padded_len // 2], padded_len // 2)
         self.prefill_buckets[padded_len // 2] = []
       if padded_len == self.max_prefill_length:

MaxText/maxengine.py

@@ -728,6 +728,10 @@ def copy(path, partial_cache, full_cache, annotations):
         zeros = jnp.zeros((1, self.config.max_prefill_predict_length), dtype=jnp.int32)
         ## zero out in case prefill cache is too small to cover
         full_cache = jax.lax.dynamic_update_index_in_dim(full_cache, zeros, slot, batch_idx)
+        # In case partial_cache is too small to slice at the given index, pad it with an extra seqlen
+        if i == num_prompts - 1:
+          pad = jnp.zeros((1, seq_len), dtype=int)
+          partial_cache = jnp.concatenate([partial_cache, pad], axis=1)
         ## copy prefill cache
         partial_cache = jax.lax.dynamic_slice(partial_cache, (0, start_idx), (1, seq_len))
         partial_cache = (partial_cache == partial_cache[0, 0]).astype(int)
@@ -748,8 +752,11 @@ def copy(path, partial_cache, full_cache, annotations):
         slice_size[seqlen_index] = seq_len
 
         slice_size = tuple(slice_size)
+        # Same as in prefill_segment_id processing
+        if i == num_prompts - 1:
+          pad = jnp.zeros(slice_size, dtype=partial_cache.dtype)
+          partial_cache = jnp.concatenate([partial_cache, pad], axis=seqlen_index)
         partial_cache = jax.lax.dynamic_slice(partial_cache, start_indices, slice_size)
-        # jax.debug.print("start_indices: {}, slice_size: {}", start_indices, slice_size)
 
         return jax.lax.dynamic_update_index_in_dim(full_cache, partial_cache, slot, batch_idx)
       else: