Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

cleanups

HenryNdubuakudbac3480c801
diff --git a/src/run.py b/src/run.py
index 5b3d798..dfb5969 100644
--- a/src/run.py
+++ b/src/run.py
@@ -85,7 +85,7 @@ def _get_decode_fn(model, max_gen_len):
def load_checkpoint(path):
with open(path, "rb") as f:
data = pickle.load(f)
- params = jax.tree.map(jnp.array, data["params"])
+ params = jax.tree.map(lambda x: jnp.array(x, dtype=jnp.bfloat16), data["params"])
config = TransformerConfig(**data["config"])
return params, config
diff --git a/src/train.py b/src/train.py
index 8dc806a..27b6d47 100644
--- a/src/train.py
+++ b/src/train.py
@@ -809,7 +809,7 @@ def train(args):
mat_results[f] = (float(math.exp(min(avg, 20))),
_estimate_mat_params(config, f), config.d_ff // f)
- params_np = jax.tree.map(np.array, eval_params)
+ params_np = jax.tree.map(lambda x: np.array(x).astype(np.float16), eval_params)
total_params = sum(x.size for x in jax.tree.leaves(params_np))
near_zero = sum(int(np.sum(np.abs(x) < 1e-6)) for x in jax.tree.leaves(params_np))
sparsity = near_zero / total_params * 100
@@ -871,9 +871,10 @@ def train(args):
unified_samples = []
for ex, pred in zip(display_pairs, display_preds):
- ref = ex["answers"].strip()
- if ref == "":
- ref = "[]"
+ try:
+ ref = _json_mod.dumps(_json_mod.loads(ex["answers"]), separators=(",", ":"))
+ except (ValueError, TypeError):
+ ref = ex["answers"].strip() or "[]"
unified_samples.append({
"query": ex["query"],
"tools": ex["tools"],