add llama3 test and is_compiling lookaside (#1397)

thunder/core/jit_ext.py

@@ -411,6 +411,20 @@ def lookaside_wrapper(lookaside):
     return lookaside_wrapper
 
 
+# PyTorch moved this to torch.compiler.is_compiling as official API
+# we are compiling
+if hasattr(torch, "compiler") and hasattr(torch.compiler, "is_compiling"):
+    is_compiling = torch.compiler.is_compiling
+else:
+    is_compiling = torch._dynamo.is_compiling
+
+
+@register_general_jit_lookaside(is_compiling)
+@interpreter_needs_wrap
+def jit_is_compiling_lookaside():
+    return True
+
+
 # lookaside for getattr. We record the provenance of the attribute but for the core attribute getting, we
 # rely on the default JIT getattr lookaside (as returned from default_lookaside)
 @register_general_jit_lookaside(getattr)

thunder/tests/test_networks.py

@@ -255,18 +255,6 @@ def test_hf_bert():
     def dummy(*args):
         pass
 
-    # Transformers 2.41+ adds some more non-essential data-dependent
-    # control flow behind a check whether we are compiling
-    if hasattr(torch, "compiler") and hasattr(torch.compiler, "is_compiling"):
-        is_compiling = torch.compiler.is_compiling
-    else:
-        is_compiling = torch._dynamo.is_compiling
-
-    @thunder.core.jit_ext.register_general_jit_lookaside(is_compiling)
-    @thunder.core.jit_ext.interpreter_needs_wrap
-    def dummy(*args):
-        return True
-
     # transformers accesses the old attrib and causes the future warning
     with warnings.catch_warnings():
         warnings.filterwarnings("ignore", category=FutureWarning, message=".*torch._dynamo.*.is_compiling.*")
@@ -410,3 +398,84 @@ def test_thunderfx_mistral_nemo_small():
     logits.backward(grad_logits)
 
     assert th_backend.subgraph_infos, "Should have at least 1 subgraph"
+
+
+LLAMA_3_2_1B_CFG = {
+    "architectures": ["LlamaForCausalLM"],
+    "attention_bias": False,
+    "attention_dropout": 0.0,
+    "bos_token_id": 128000,
+    "eos_token_id": 128001,
+    "head_dim": 64,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "intermediate_size": 8192,
+    "max_position_embeddings": 131072,
+    "mlp_bias": False,
+    "model_type": "llama",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 16,
+    "num_key_value_heads": 8,
+    "pretraining_tp": 1,
+    "rms_norm_eps": 1e-05,
+    "rope_scaling": {
+        "factor": 32.0,
+        "high_freq_factor": 4.0,
+        "low_freq_factor": 1.0,
+        "original_max_position_embeddings": 8192,
+        "rope_type": "llama3",
+    },
+    "rope_theta": 500000.0,
+    "tie_word_embeddings": True,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.45.0.dev0",
+    "use_cache": True,
+    "vocab_size": 128256,
+    "_commit_hash": "4e20de362430cd3b72f300e6b0f18e50e7166e08",
+}
+
+
+@requiresCUDA
+def test_hf_llama():
+    from transformers.models.llama import LlamaForCausalLM, LlamaConfig
+    from transformers import DynamicCache
+    from transformers.models.llama.modeling_llama import logger as llama_logger
+    import logging
+
+    # transformers logs a cache deprecation warning
+    llama_logger.setLevel(logging.CRITICAL)
+    model_id = "meta-llama/Llama-3.2-1B"
+
+    config_args = LLAMA_3_2_1B_CFG.copy()
+    config_args["num_hidden_layers"] = 1
+    with torch.device("cuda"):
+        model = LlamaForCausalLM(LlamaConfig(**config_args)).to(torch.bfloat16).requires_grad_(False).eval()
+
+    jm = thunder.jit(model)
+
+    args1 = dict(
+        cache_position=torch.tensor([0, 1, 2, 3, 4, 5], device="cuda:0"),
+        input_ids=torch.tensor([[128000, 791, 1401, 311, 2324, 374]], device="cuda:0"),
+        inputs_embeds=None,
+        attention_mask=torch.tensor([[1, 1, 1, 1, 1, 1]], device="cuda:0"),
+        use_cache=True,
+        return_dict=True,
+    )
+    res = jm(**args1)
+    expected = model(**args1)
+
+    assert_close(res, expected, rtol=1e-1, atol=1e-1)
+
+    args2 = dict(
+        cache_position=torch.tensor([6], device="cuda:0"),
+        input_ids=torch.tensor([[311]], device="cuda:0"),
+        inputs_embeds=None,
+        attention_mask=torch.tensor([[1, 1, 1, 1, 1, 1, 1]], device="cuda:0"),
+        use_cache=True,
+        return_dict=True,
+    )
+
+    res2 = jm(past_key_values=res["past_key_values"], **args2)
+    expected2 = model(past_key_values=res["past_key_values"], **args2)
+    assert_close(res2, expected2, rtol=1e-1, atol=1e-1)