Port #1095 to skyrl folder (#1129)

See #1095
<!-- devin-review-badge-begin -->

---

<a href="https://app.devin.ai/review/novasky-ai/skyrl/pull/1129"
target="_blank">
  <picture>
<source media="(prefers-color-scheme: dark)"
srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img
src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1"
alt="Open with Devin">
  </picture>
</a>
<!-- devin-review-badge-end -->

Author: pcmoritz

skyrl/tests/tx/models/conftest.py

@@ -0,0 +1,30 @@
+from flax import nnx
+import jax
+import jax.numpy as jnp
+from transformers import AutoConfig
+
+from skyrl.tx.models.configs import ModelConfig
+from skyrl.tx.models.types import ModelForCausalLM
+from skyrl.tx.utils.models import load_safetensors, resolve_model_path
+
+
+def load_model(
+    model_name: str,
+    config_cls: type[ModelConfig],
+    model_cls: type[ModelForCausalLM],
+    mesh_axes: tuple[str, ...],
+    *,
+    mesh_shape: tuple[int, ...] | None = None,
+    **config_kwargs,
+) -> tuple[ModelConfig, ModelForCausalLM]:
+    """Create a JAX model and load weights from the HuggingFace cache."""
+    weights_dir = resolve_model_path(model_name)
+    base_config = AutoConfig.from_pretrained(model_name)
+    config = config_cls(base_config, shard_attention_heads=True, **config_kwargs)
+    if mesh_shape is None:
+        mesh_shape = (1,) * len(mesh_axes)
+    mesh = jax.make_mesh(mesh_shape, mesh_axes, axis_types=(jax.sharding.AxisType.Auto,) * len(mesh_axes))
+    with jax.set_mesh(mesh):
+        model = model_cls(config, dtype=jnp.float32, rngs=nnx.Rngs(0))
+    load_safetensors(weights_dir, config, model)
+    return config, model

skyrl/tests/tx/models/test_models_common.py

@@ -1,17 +1,14 @@
-import tempfile
-
-from flax import nnx
-import jax
 import jax.numpy as jnp
 import numpy as np
 import pytest
-from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from skyrl.tx.models.configs import Llama3Config, ModelConfig, Qwen3Config
 from skyrl.tx.models.llama3 import Llama3ForCausalLM
 from skyrl.tx.models.qwen3 import Qwen3ForCausalLM
 from skyrl.tx.models.types import ModelForCausalLM
-from skyrl.tx.utils.models import load_safetensors
+
+from tests.tx.models.conftest import load_model
 
 MODEL_PARAMS = [
     ("unsloth/Llama-3.2-1B", Llama3Config, Llama3ForCausalLM, ("fsdp", "tp")),
@@ -20,32 +17,6 @@
 MODEL_IDS = ["llama3", "qwen3"]
 
 
-def load_model(
-    tmp_dir: str,
-    model_name: str,
-    config_cls: type[ModelConfig],
-    model_cls: type[ModelForCausalLM],
-    mesh_axes: tuple[str, str],
-    *,
-    loss_chunk_size: int = 0,
-) -> ModelForCausalLM:
-    """Load model from pre-saved weights directory."""
-    base_config = AutoConfig.from_pretrained(model_name)
-    config = config_cls(
-        base_config,
-        max_lora_adapters=1,
-        max_lora_rank=1,
-        shard_attention_heads=True,
-        loss_chunk_size=loss_chunk_size,
-        gradient_checkpointing=False,
-    )
-    mesh = jax.make_mesh((1, 1), mesh_axes, axis_types=(jax.sharding.AxisType.Auto,) * 2)
-    with jax.set_mesh(mesh):
-        model = model_cls(config, dtype=jnp.float32, rngs=nnx.Rngs(0))
-    load_safetensors(tmp_dir, config, model)
-    return model
-
-
 @pytest.mark.parametrize("model_name,config_cls,model_cls,mesh_axes", MODEL_PARAMS, ids=MODEL_IDS)
 def test_compute_logits(
     model_name: str,
@@ -59,22 +30,27 @@ def test_compute_logits(
     inputs = ["The capital of France is", "Hello world"]
     batch = tokenizer(inputs, return_tensors="pt", padding=True)
 
-    with tempfile.TemporaryDirectory() as tmp:
-        # Load HF model, get logits, save weights, then delete to free memory
-        hf_model = AutoModelForCausalLM.from_pretrained(model_name, attn_implementation="eager", use_safetensors=True)
-        hf_outputs = hf_model(batch.input_ids, attention_mask=batch.attention_mask)
-        hf_logits = hf_outputs.logits.detach().numpy()
-        hf_model.save_pretrained(tmp, safe_serialization=True)
-        del hf_model, hf_outputs
-
-        # Load our model from saved weights
-        model = load_model(tmp, model_name, config_cls, model_cls, mesh_axes)
+    # Load HF model, get logits, then delete to free memory
+    hf_model = AutoModelForCausalLM.from_pretrained(model_name, attn_implementation="eager", use_safetensors=True)
+    hf_outputs = hf_model(batch.input_ids, attention_mask=batch.attention_mask)
+    hf_logits = hf_outputs.logits.detach().numpy()
+    del hf_model, hf_outputs
+
+    _, model = load_model(
+        model_name,
+        config_cls,
+        model_cls,
+        mesh_axes,
+        max_lora_adapters=1,
+        max_lora_rank=1,
+        gradient_checkpointing=False,
+    )
 
-        # Get our logits via compute_logits
-        outputs = model(batch.input_ids.numpy(), attention_mask=batch.attention_mask.numpy())
-        our_logits = np.asarray(model.compute_logits(outputs.last_hidden_state))
+    # Get our logits via compute_logits
+    outputs = model(batch.input_ids.numpy(), attention_mask=batch.attention_mask.numpy())
+    our_logits = np.asarray(model.compute_logits(outputs.last_hidden_state))
 
-        np.testing.assert_allclose(our_logits, hf_logits, rtol=3e-2, atol=3e-2)
+    np.testing.assert_allclose(our_logits, hf_logits, rtol=3e-2, atol=3e-2)
 
 
 @pytest.mark.parametrize("model_name,config_cls,model_cls,mesh_axes", MODEL_PARAMS, ids=MODEL_IDS)
@@ -94,22 +70,18 @@ def test_chunked_logprobs(
     attention_mask = jnp.array(batch.attention_mask.numpy())
     target_ids = jnp.roll(input_ids, -1, axis=1)
 
-    with tempfile.TemporaryDirectory() as tmp:
-        # Save HF weights once
-        hf_model = AutoModelForCausalLM.from_pretrained(model_name, attn_implementation="eager", use_safetensors=True)
-        hf_model.save_pretrained(tmp, safe_serialization=True)
-        del hf_model
-
-        # Load non-chunked model, compute logprobs, then delete
-        model = load_model(tmp, model_name, config_cls, model_cls, mesh_axes, loss_chunk_size=0)
-        outputs = model(input_ids, attention_mask=attention_mask)
-        logprobs_nonchunked = np.asarray(model.compute_logprobs(outputs.last_hidden_state, target_ids))
-        del model, outputs
-
-        # Load chunked model, compute logprobs
-        model = load_model(tmp, model_name, config_cls, model_cls, mesh_axes, loss_chunk_size=chunk_size)
-        outputs = model(input_ids, attention_mask=attention_mask)
-        logprobs_chunked = np.asarray(model.compute_logprobs(outputs.last_hidden_state, target_ids))
+    common_kwargs = dict(max_lora_adapters=1, max_lora_rank=1, gradient_checkpointing=False)
+
+    # Load non-chunked model, compute logprobs, then delete
+    _, model = load_model(model_name, config_cls, model_cls, mesh_axes, loss_chunk_size=0, **common_kwargs)
+    outputs = model(input_ids, attention_mask=attention_mask)
+    logprobs_nonchunked = np.asarray(model.compute_logprobs(outputs.last_hidden_state, target_ids))
+    del model, outputs
+
+    # Load chunked model, compute logprobs
+    _, model = load_model(model_name, config_cls, model_cls, mesh_axes, loss_chunk_size=chunk_size, **common_kwargs)
+    outputs = model(input_ids, attention_mask=attention_mask)
+    logprobs_chunked = np.asarray(model.compute_logprobs(outputs.last_hidden_state, target_ids))
 
     np.testing.assert_allclose(
         logprobs_chunked,