OLMo 2 (WIP)

litgpt/config.py

@@ -78,6 +78,7 @@ class Config:
     scale_embeddings: bool = False
     lm_head_bias: bool = False
     final_logit_softcapping: Optional[float] = None
+    input_norm: bool = True
 
     def __post_init__(self):
         if not self.name:
@@ -912,6 +913,62 @@ def norm_class(self) -> Type:
 
 configs.extend(olmo)
 
+olmo2 = [
+    # https://huggingface.co/allenai/OLMo-2-1124-7B/blob/main/config.json
+    dict(
+        name="OLMo-2-1124-7B{}",
+        hf_config=dict(org="allenai", name="OLMo-2-1124-7B{}"),
+        vocab_size=100278,
+        padded_vocab_size=100352,
+        block_size=4096,
+        n_embd=4096,
+        n_layer=32,
+        n_head=32,
+        n_query_groups=32,
+        rotary_percentage=1.0,
+        parallel_residual=False,
+        bias=False,
+        norm_class_name="RMSNorm",
+        mlp_class_name="LLaMAMLP",
+        norm_eps=1e-06,
+        intermediate_size=11008,
+        rope_base=500000,
+        norm_qk=True,
+        post_mlp_norm=True,
+        input_norm=False,
+    ),
+    # https://huggingface.co/allenai/OLMo-2-1124-13B/blob/main/config.json
+    dict(
+        name="OLMo-2-1124-13B{}",
+        hf_config=dict(org="allenai", name="OLMo-2-1124-13B{}"),
+        vocab_size=100278,
+        padded_vocab_size=100352,
+        block_size=4096,
+        n_embd=5120,
+        n_layer=40,
+        n_head=40,
+        n_query_groups=40,
+        rotary_percentage=1.0,
+        parallel_residual=False,
+        bias=False,
+        norm_class_name="RMSNorm",
+        mlp_class_name="LLaMAMLP",
+        norm_eps=1e-06,
+        intermediate_size=13824,
+        rope_base=500000,
+        norm_qk=True,
+        post_mlp_norm=True,
+        input_norm=False,
+    ),
+]
+
+for c in olmo2:
+    for kind in ("", "-SFT", "-DPO", "-Instruct"):
+        copy = deepcopy(c)
+        copy["name"] = c["name"].format(kind)
+        copy["hf_config"]["name"] = c["hf_config"]["name"].format(kind)
+        configs.append(copy)
+
 ###############
 # Google Gemma
 ###############

litgpt/model.py

@@ -263,7 +263,7 @@ def __init__(
                 " (non-parallel residual and shared attention norm)."
             )
 
-        self.norm_1 = config.norm_class(config.n_embd, eps=config.norm_eps)
+        self.norm_1 = None if not config.input_norm else config.norm_class(config.n_embd, eps=config.norm_eps)
         self.attn = CausalSelfAttention(config, block_idx)
         self.post_attention_norm = (
             config.norm_class(config.n_embd, eps=config.norm_eps) if config.post_attention_norm else nn.Identity()
@@ -306,7 +306,11 @@ def forward(
         └───► +
         """
 
-        x_normed = self.norm_1(x)
+        if self.norm_1 is not None:
+            x_normed = self.norm_1(x)
+        else:
+            x_normed = x
+
         attention_output = self.attn(
             x_normed, cos, sin, mask, input_pos, input_pos_maxp1
         )
@@ -342,11 +346,12 @@ def __init__(self, config: Config, block_idx: int) -> None:
             block_idx % config.sliding_window_layer_stride == 0
         )
 
+
+        self.q_norm = None
+        self.k_norm = None
         if config.norm_qk:
-            self.norm_q = config.norm_class(config.head_size * config.n_head, eps=config.norm_eps)
-            self.norm_k = config.norm_class(config.head_size * config.n_query_groups, eps=config.norm_eps)
-        else:
-            self.norm_q = self.norm_k = None
+            self.q_norm = config.norm_class(config.head_size * config.n_head, eps=config.norm_eps)
+            self.k_norm = config.norm_class(config.head_size * config.n_query_groups, eps=config.norm_eps)
 
         self.config = config
         self.block_idx = block_idx
@@ -385,8 +390,8 @@ def forward(
         q, k, v = qkv.split((query_size, key_size, value_size), dim=-1)  # 3x(B, T, C*)
 
         if self.config.norm_qk:
-            q = self.norm_q(q)
-            k = self.norm_k(k)
+            q = self.q_norm(q)
+            k = self.k_norm(k)
 
         # To place the num_heads (nh) dimension right after the batch (B) dimension, the first step is to decouple the
         # embedding size (C) into num_heads (nh) and head_size (hs).

litgpt/prompts.py

@@ -368,6 +368,8 @@ def model_name_to_prompt_style(model_name: str) -> PromptStyle:
         return Llama3()
     if re.search("Llama-3.*-Instruct-*", model_name):
         return Llama3()
+    if re.search("OLMo-2.*-(Instruct|SFT|DPO)", model_name):
+        return Llama3()
     if re.search("FreeWilly2", model_name):
         return FreeWilly2()
     if re.search("Platypus", model_name):

litgpt/scripts/convert_hf_checkpoint.py

@@ -447,6 +447,85 @@ def copy_weights_qwen_2_5(
                 pbar.update(progress_per_file)
 
 
+def copy_weights_olmo2(
+    config: Config,
+    qkv_weights: Dict[int, List[Optional[NotYetLoadedTensor]]],
+    state_dict: Dict[str, torch.Tensor],
+    hf_weights: Dict[str, Union[torch.Tensor, NotYetLoadedTensor]],
+    saver: Optional[incremental_save] = None,
+    dtype: Optional[torch.dtype] = None,
+    pbar: Optional[tqdm] = None,
+    progress_per_file: Optional[float] = None,
+    debug_mode: Optional[bool] = False,
+) -> None:
+    weight_map = {
+        "model.embed_tokens.weight": "transformer.wte.weight",
+        "model.layers.{}.self_attn.q_norm.weight": "transformer.h.{}.attn.q_norm.weight",
+        "model.layers.{}.self_attn.q_proj.weight": None,
+        "model.layers.{}.self_attn.k_norm.weight": "transformer.h.{}.attn.k_norm.weight",
+        "model.layers.{}.self_attn.k_proj.weight": None,
+        "model.layers.{}.self_attn.v_proj.weight": None,
+        "model.layers.{}.self_attn.o_proj.weight": "transformer.h.{}.attn.proj.weight",
+        "model.layers.{}.self_attn.rotary_emb.inv_freq": None,
+        "model.layers.{}.post_attention_layernorm.weight": "transformer.h.{}.norm_2.weight",
+        "model.layers.{}.post_attention_layernorm.bias": "transformer.h.{}.norm_2.bias",
+        "model.layers.{}.post_feedforward_layernorm.weight": "transformer.h.{}.post_mlp_norm.weight",
+        "model.norm.weight": "transformer.ln_f.weight",
+        "model.norm.bias": "transformer.ln_f.bias",
+        "lm_head.weight": "lm_head.weight",
+    }
+    if config.mlp_class_name in ("LLaMAMLP", "GemmaMLP"):
+        weight_map.update(
+            {
+                "model.layers.{}.mlp.gate_proj.weight": "transformer.h.{}.mlp.fc_1.weight",
+                "model.layers.{}.mlp.up_proj.weight": "transformer.h.{}.mlp.fc_2.weight",
+                "model.layers.{}.mlp.down_proj.weight": "transformer.h.{}.mlp.proj.weight",
+            }
+        )
+    else:
+        raise NotImplementedError
+
+    if progress_per_file is not None:
+        progress_per_file = progress_per_file / max(1, len(hf_weights) + len(qkv_weights))
+
+    for from_name, param in hf_weights.items():
+        name_template, *ids = layer_template(from_name, num_matches=2)
+        to_name = weight_map[name_template]
+        param = load_param(param, from_name, dtype, verbose=debug_mode)
+        if any(w in from_name for w in ("q_proj", "k_proj", "v_proj")):
+            qkv = qkv_weights.setdefault(ids[0], defaultdict(dict))
+            weight_name, weight_type = from_name.split(".")[-2:]
+            qkv[weight_type][weight_name] = param
+        if to_name is None:
+            continue
+        to_name = to_name.format(*ids)
+        if saver is not None:
+            param = saver.store_early(param)
+        state_dict[to_name] = param
+
+        if progress_per_file is not None:
+            pbar.update(progress_per_file)
+
+    if "lm_head.weight" not in state_dict:
+        state_dict["lm_head.weight"] = state_dict["transformer.wte.weight"]
+
+    for i in list(qkv_weights):
+        for weight_type in list(qkv_weights[i]):
+            qkv = qkv_weights[i][weight_type]
+            if len(qkv) != 3:
+                # qkv is splitted across different .bin files
+                continue
+            q = load_param(qkv["q_proj"], f"layer {i} q {weight_type}", dtype, verbose=debug_mode)
+            k = load_param(qkv["k_proj"], f"layer {i} k {weight_type}", dtype, verbose=debug_mode)
+            v = load_param(qkv["v_proj"], f"layer {i} v {weight_type}", dtype, verbose=debug_mode)
+            qkv = torch.cat((q, k, v))
+            state_dict[f"transformer.h.{i}.attn.qkv.{weight_type}"] = qkv
+            del qkv_weights[i][weight_type]
+
+            if progress_per_file is not None:
+                pbar.update(progress_per_file)
+
+
 def qkv_reassemble(
     param: Union[torch.Tensor, NotYetLoadedTensor], config: Config
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
@@ -537,6 +616,10 @@ def convert_hf_checkpoint(
         # holder to reconstitute the split q, k, v
         qkv_weights = {}
         copy_fn = partial(copy_weights_qwen_2_5, config, qkv_weights)
+    elif model_name.lower().startswith(("olmo-2-")):
+        # holder to reconstitute the split q, k, v
+        qkv_weights = {}
+        copy_fn = partial(copy_weights_olmo2, config, qkv_weights)
     elif config.mlp_class_name in ("LLaMAMLP", "GemmaMLP", "LLaMAMoE"):
         # holder to reconstitute the split q, k, v
         qkv_weights = {}

litgpt/scripts/convert_lit_checkpoint.py

@@ -339,6 +339,64 @@ def copy_weights_qwen_2_5(
             state_dict[to_name] = param
 
 
+def copy_weights_olmo2(
+    config: Config,
+    state_dict: Dict[str, torch.Tensor],
+    lit_weights: Dict[str, Union[torch.Tensor, NotYetLoadedTensor]],
+    untie_weights: bool = False,
+    saver: Optional[incremental_save] = None,
+) -> None:
+    weight_map = {
+        "transformer.wte.weight": "model.embed_tokens.weight",
+        "transformer.h.{}.attn.proj.weight": "model.layers.{}.self_attn.o_proj.weight",
+        "transformer.h.{}.attn.q_norm.weight": "model.layers.{}.self_attn.q_norm.weight",
+        "transformer.h.{}.attn.k_norm.weight": "model.layers.{}.self_attn.k_norm.weight",
+        "transformer.h.{}.norm_2.weight": "model.layers.{}.post_attention_layernorm.weight",
+        "transformer.h.{}.norm_2.bias": "model.layers.{}.post_attention_layernorm.bias",
+        "transformer.h.{}.post_mlp_norm.weight": "model.layers.{}.post_feedforward_layernorm.weight",
+        "transformer.ln_f.weight": "model.norm.weight",
+        "transformer.ln_f.bias": "model.norm.bias",
+        "lm_head.weight": "lm_head.weight",
+    }
+    if config.mlp_class_name in ("LLaMAMLP", "GemmaMLP"):
+        weight_map.update(
+            {
+                "transformer.h.{}.mlp.fc_1.weight": "model.layers.{}.mlp.gate_proj.weight",
+                "transformer.h.{}.mlp.fc_2.weight": "model.layers.{}.mlp.up_proj.weight",
+                "transformer.h.{}.mlp.proj.weight": "model.layers.{}.mlp.down_proj.weight",
+            }
+        )
+    else:
+        raise NotImplementedError
+
+    for from_name, param in lit_weights.items():
+        if from_name == "lm_head.weight" and untie_weights:
+            continue
+        name_template, *ids = layer_template(from_name, num_matches=2)
+        param = load_param(param, from_name, None)
+        if from_name.endswith(".attn.qkv.weight"):
+            to_names = (
+                "model.layers.{}.self_attn.q_proj.weight".format(*ids),
+                "model.layers.{}.self_attn.k_proj.weight".format(*ids),
+                "model.layers.{}.self_attn.v_proj.weight".format(*ids),
+            )
+            params = param.split(
+                (
+                    config.n_head * config.head_size,
+                    config.n_query_groups * config.head_size,
+                    config.n_query_groups * config.head_size,
+                )
+            )
+        else:
+            to_names = (weight_map[name_template].format(*ids),)
+            params = (param,)
+
+        for to_name, param in zip(to_names, params):
+            if saver is not None:
+                param = saver.store_early(param)
+            state_dict[to_name] = param
+
+
 def qkv_reassemble(param: Union[torch.Tensor, NotYetLoadedTensor], config: Config) -> torch.Tensor:
     """Reassemble from a normal to an interleaved placement in a QKV matrix.
     [Q, Q, ..., K, K, ..., V, V, ...] --> [Q, K, V, Q, K, V, ...]
@@ -383,6 +441,8 @@ def convert_lit_checkpoint(checkpoint_dir: Path, output_dir: Path) -> None:
         copy_fn = partial(copy_weights_phi, config)
     elif config.name.lower().startswith(("qwen2.5","qwq")):
         copy_fn = partial(copy_weights_qwen_2_5, config)
+    elif config.name.lower().startswith(("olmo-2-")):
+        copy_fn = partial(copy_weights_olmo2, config)
     elif config.mlp_class_name in ("LLaMAMLP", "GemmaMLP", "LLaMAMoE"):
         untie_weights = "Gemma" in config.name
         copy_fn = partial(copy_weights_llama, config, untie_weights=untie_weights)

tests/test_model.py

@@ -28,6 +28,7 @@
 from transformers.models.mistral import MistralConfig, MistralForCausalLM
 from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
 from transformers.models.olmo import OlmoConfig, OlmoForCausalLM
+from transformers.models.olmo2 import Olmo2Config, Olmo2ForCausalLM
 from transformers.models.qwen2 import Qwen2Config, Qwen2ForCausalLM
 
 import litgpt.config as config_module
@@ -38,6 +39,7 @@
     copy_weights_gemma_2,
     copy_weights_gpt_neox,
     copy_weights_hf_llama,
+    copy_weights_olmo2,
     copy_weights_phi,
     copy_weights_qwen_2_5,
 )
@@ -617,6 +619,65 @@ def test_against_olmo(model_name, device, dtype):
     torch.testing.assert_close(ours_y, theirs_y)
 
 
+@torch.inference_mode()
+@pytest.mark.parametrize("model_name", ("OLMo-2-1124-7B", "OLMo-2-1124-13B"))
+@pytest.mark.parametrize(
+    ("device", "dtype"),
+    [
+        (torch.device("cpu"), torch.float32),
+        pytest.param(
+            torch.device("cuda"),
+            torch.float16,
+            marks=[
+                # the reference does softmax upscaled to fp32 during attention. additionally, the final layernorm input
+                # is slightly different
+                pytest.mark.xfail(raises=AssertionError, strict=False),
+                RunIf(min_cuda_gpus=1),
+            ],
+        ),
+    ],
+)
+def test_against_olmo2(model_name, device, dtype):
+    torch.set_default_dtype(dtype)
+
+    ours_config = Config.from_name(
+        model_name,
+        padded_vocab_size=10000,
+        n_layer=2,
+        n_head=8,
+        n_embd=32,
+        n_query_groups=2,
+        intermediate_size=86,
+    )
+    T = 5
+    theirs_config = Olmo2Config(
+        vocab_size=ours_config.padded_vocab_size,
+        hidden_size=ours_config.n_embd,
+        intermediate_size=ours_config.intermediate_size,
+        num_hidden_layers=ours_config.n_layer,
+        num_attention_heads=ours_config.n_head,
+        num_key_value_heads=ours_config.n_query_groups,
+        max_positional_embeddings=T,
+        attention_bias=ours_config.bias,
+        rope_theta=ours_config.rope_base,
+    )
+    assert ours_config.intermediate_size == theirs_config.intermediate_size
+
+    theirs_model = Olmo2ForCausalLM(theirs_config).to(device)
+    theirs_state_dict = theirs_model.state_dict()
+    state_dict = {}
+    copy_weights_olmo2(ours_config, {}, state_dict, theirs_state_dict)
+    ours_model = GPT(ours_config).to(device)
+    ours_model.load_state_dict(state_dict)
+
+    # test end to end
+    x = torch.tensor([[9856, 23, 491, 1536, 304]], dtype=torch.int32, device=device)
+    assert x.size(1) == T
+    ours_y = ours_model(x)
+    theirs_y = theirs_model(x)["logits"].to(dtype)  # HF converts logits to float
+    torch.testing.assert_close(ours_y, theirs_y)
+
+
 @torch.inference_mode()
 @pytest.mark.parametrize(
     ("device", "dtype"),