Llama 3.1 RoPE scaling (#205)

* feat(llama): import RoPE scaling code

This is imported from the original Llama reference implementation:
https://github.com/meta-llama/llama-models/blob/7890266c5a3ccd29e739d53a71ea968bcf4ca400/models/llama3/reference_impl/model.py#L45

Note that the function does not have any effect on the original model
code as long as the use_scaled parameter is false (the default).

* feat(llama): add RopeScalingArgs

These are aligned with HF ones, so it will be easier to implement rope
scaling as it is done in Llama3.1.

* feat(llama): support rope scaling arguments to improve flexibility

* chore: relax safetensors pattern on download

* feat: untie weights when needed (i.e.: Llama3.2-1B)

* feat: add support for Llama3.1 - 3.2 and 3.3 models

README.md

@@ -67,6 +67,12 @@ meta-llama/Meta-Llama-3-8B
 meta-llama/Meta-Llama-3-8B-Instruct
 meta-llama/Meta-Llama-3-70B
 meta-llama/Meta-Llama-3-70B-Instruct
+meta-llama/Llama-3.1-8B
+meta-llama/Llama-3.1-8B-Instruct
+meta-llama/Llama-3.2-1B
+meta-llama/Llama-3.2-1B-Instruct
+meta-llama/Llama-3.3-70B
+meta-llama/Llama-3.3-70B-Instruct
 google/gemma-2b
 google/gemma-2b-it
 google/gemma-7b

jetstream_pt/cli.py

@@ -44,6 +44,9 @@
 def shard_weights(env, weights, weight_shardings):
   """Shard weights according to weight_shardings"""
   sharded = {}
+  # Some output and embeddings weights might be tied: in this case untie them
+  if weights["output.weight"].device.type == "meta":
+    weights["output.weight"] = weights["tok_embeddings.weight"].clone()
   for key, val in weights.items():
     sharding = env.sharding_by_axis(weight_shardings.get(key, -1))
     with jax.default_device(jax.devices("cpu")[0]):

jetstream_pt/fetch_models.py

@@ -60,6 +60,9 @@ class ModelInfo:
 _llama2_70 = ModelInfo(llama_model.Transformer, 80, 8, 128, 8)
 _llama3_8 = ModelInfo(llama_model.Transformer, 32, 8, 128, 4)
 _llama3_70 = _llama2_70
+_llama3_1_8b = _llama3_8
+_llama3_2_1b = ModelInfo(llama_model.Transformer, 16, 8, 64, 4)
+_llama3_3_70b = _llama2_70
 
 _mixtral_87 = ModelInfo(mixtral_model.Transformer, 32, 8, 128, 4)
 
@@ -78,6 +81,12 @@ class ModelInfo:
     "meta-llama/Meta-Llama-3-8B-Instruct": _llama3_8,
     "meta-llama/Meta-Llama-3-70B": _llama3_70,
     "meta-llama/Meta-Llama-3-70B-Instruct": _llama3_70,
+    "meta-llama/Llama-3.1-8B": _llama3_1_8b,
+    "meta-llama/Llama-3.1-8B-Instruct": _llama3_1_8b,
+    "meta-llama/Llama-3.2-1B": _llama3_2_1b,
+    "meta-llama/Llama-3.2-1B-Instruct": _llama3_2_1b,
+    "meta-llama/Llama-3.3-70B": _llama3_3_70b,
+    "meta-llama/Llama-3.3-70B-Instruct": _llama3_3_70b,
     "google/gemma-2b": _gemma_2b,
     "google/gemma-2b-it": _gemma_2b,
     "google/gemma-7b": _gemma_7b,
@@ -215,7 +224,7 @@ def _hf_download(
         local_dir_use_symlinks=False,
         token=hf_token,
         allow_patterns=[
-            "model-?????-of-?????.safetensors",
+            "model*.safetensors",
             "*.json",
             "*.model",
         ],

jetstream_pt/third_party/llama/model_args.py

@@ -5,6 +5,16 @@
 from typing import Optional
 
 
+@dataclasses.dataclass
+class RopeScalingArgs:
+  """Rope scaling configuration parameters."""
+
+  factor: float = 8.0
+  low_freq_factor: float = 1.0
+  high_freq_factor: float = 4.0
+  original_max_position_embeddings: int = 8192
+
+
 @dataclasses.dataclass
 class ModelArgs:
   """Model configuration parameters."""
@@ -29,6 +39,7 @@ class ModelArgs:
   device = "cpu"
 
   rope_theta: float = 10000.0
+  rope_scaling_args: RopeScalingArgs = None
 
 
 def get_arg(
@@ -103,6 +114,60 @@ def get_arg(
         "vocab_size": 128256,
         "rope_theta": 500000.0,
     }
+  elif model_name == "llama-3.1-8b":
+    data = {
+        "dim": 4096,
+        "vocab_size": 128256,
+        "multiple_of": 1024,
+        "ffn_dim_multiplier": 1.3,
+        "n_layers": 32,
+        "n_heads": 32,
+        "n_kv_heads": 8,
+        "norm_eps": 1e-05,
+        "rope_theta": 500000.0,
+        "rope_scaling_args": RopeScalingArgs(
+            factor=8.0,
+            low_freq_factor=1.0,
+            high_freq_factor=4.0,
+            original_max_position_embeddings=8192,
+        ),
+    }
+  elif model_name == "llama-3.2-1b":
+    data = {
+        "dim": 2048,
+        "vocab_size": 128256,
+        "multiple_of": 1024,
+        "ffn_dim_multiplier": 1.5,
+        "n_layers": 16,
+        "n_heads": 32,
+        "n_kv_heads": 8,
+        "norm_eps": 1e-05,
+        "rope_theta": 500000.0,
+        "rope_scaling_args": RopeScalingArgs(
+            factor=32.0,
+            low_freq_factor=1.0,
+            high_freq_factor=4.0,
+            original_max_position_embeddings=8192,
+        ),
+    }
+  elif model_name == "llama-3.3-70b":
+    data = {
+        "dim": 8192,
+        "vocab_size": 128256,
+        "multiple_of": 1024,
+        "ffn_dim_multiplier": 1.3,
+        "n_layers": 80,
+        "n_heads": 64,
+        "n_kv_heads": 8,
+        "norm_eps": 1e-05,
+        "rope_theta": 500000.0,
+        "rope_scaling_args": RopeScalingArgs(
+            factor=8.0,
+            low_freq_factor=1.0,
+            high_freq_factor=4.0,
+            original_max_position_embeddings=8192,
+        ),
+    }
 
   return ModelArgs(
       max_seq_len=seqlen,

jetstream_pt/third_party/llama/model_exportable.py

@@ -4,6 +4,7 @@
 from typing import Any, List, Optional
 import copy
 import jax
+import math
 import torch
 import torch.nn.functional as F
 import functools
@@ -170,12 +171,42 @@ def forward(
       return out
 
 
+def apply_scaling(freqs: torch.Tensor, config: model_args.RopeScalingArgs):
+  # Values obtained from grid search
+  scale_factor = config.factor
+  low_freq_factor = config.low_freq_factor
+  high_freq_factor = config.high_freq_factor
+  old_context_len = config.original_max_position_embeddings
+
+  low_freq_wavelen = old_context_len / low_freq_factor
+  high_freq_wavelen = old_context_len / high_freq_factor
+  new_freqs = []
+  for freq in freqs:
+    wavelen = 2 * math.pi / freq
+    if wavelen < high_freq_wavelen:
+      new_freqs.append(freq)
+    elif wavelen > low_freq_wavelen:
+      new_freqs.append(freq / scale_factor)
+    else:
+      assert low_freq_wavelen != high_freq_wavelen
+      smooth = (old_context_len / wavelen - low_freq_factor) / (
+          high_freq_factor - low_freq_factor
+      )
+      new_freqs.append((1 - smooth) * freq / scale_factor + smooth * freq)
+  return torch.tensor(new_freqs, dtype=freqs.dtype, device=freqs.device)
+
+
 def precompute_freqs_cis(
-    dim: int, end: int, theta: float = 10000.0
-) -> torch.Tensor:
+    dim: int,
+    end: int,
+    theta: float = 10000.0,
+    rope_scaling_config: model_args.RopeScalingArgs = None,
+):
   freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
-  t = torch.arange(end, device=freqs.device)  # type: ignore
-  freqs = torch.outer(t, freqs).float()  # type: ignore
+  t = torch.arange(end, device=freqs.device, dtype=torch.float32)
+  if rope_scaling_config is not None:
+    freqs = apply_scaling(freqs, rope_scaling_config)
+  freqs = torch.outer(t, freqs)
   freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
   return freqs_cis
 
@@ -223,6 +254,7 @@ def __init__(
         self.params.dim // self.params.n_heads,
         self.params.max_seq_len * 2,
         theta=self.params.rope_theta,
+        rope_scaling_config=self.params.rope_scaling_args,
     )
 
     self.register_buffer("freqs_cis", freqs_cis)
@@ -306,6 +338,12 @@ def from_hf_model_id(cls, model_id, env, is_tiny=False):
           "meta-llama/Meta-Llama-3-8B-Instruct": "llama-3-8b",
           "meta-llama/Meta-Llama-3-70B": "llama-3-70b",
           "meta-llama/Meta-Llama-3-70B-Instruct": "llama-3-70b",
+          "meta-llama/Llama-3.1-8B": "llama-3.1-8b",
+          "meta-llama/Llama-3.1-8B-Instruct": "llama-3.1-8b",
+          "meta-llama/Llama-3.2-1B": "llama-3.2-1b",
+          "meta-llama/Llama-3.2-1B-Instruct": "llama-3.2-1b",
+          "meta-llama/Llama-3.3-70B": "llama-3.3-70b",
+          "meta-llama/Llama-3.3-70B-Instruct": "llama-3.3-70b",
       }.get(model_id)
     assert name
     args = model_args.get_model_args(