Llama-2-70b-hf / llama_updates.patch
osanseviero's picture
Squashing commit
a5dda7c
diff --git a/src/transformers/models/llama/configuration_llama.py b/src/transformers/models/llama/configuration_llama.py
index d456b79e6..f85603289 100644
--- a/src/transformers/models/llama/configuration_llama.py
+++ b/src/transformers/models/llama/configuration_llama.py
@@ -50,6 +50,9 @@ class LlamaConfig(PretrainedConfig):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer encoder.
+ num_key_value_heads (`int`, *optional*, defaults to 32):
+ This is the number of groups that should be used to implement GQA.When converting a multi-head checkpoint to a GQA checkpoint, we
+ construct each group key and value head by meanpooling all the original heads within that group
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to 2048):
@@ -97,6 +100,7 @@ class LlamaConfig(PretrainedConfig):
intermediate_size=11008,
num_hidden_layers=32,
num_attention_heads=32,
+ num_key_value_heads=32,
hidden_act="silu",
max_position_embeddings=2048,
initializer_range=0.02,
@@ -115,6 +119,7 @@ class LlamaConfig(PretrainedConfig):
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
+ self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
diff --git a/src/transformers/models/llama/convert_llama_weights_to_hf.py b/src/transformers/models/llama/convert_llama_weights_to_hf.py
index e8fb7f825..a9464e1a6 100644
--- a/src/transformers/models/llama/convert_llama_weights_to_hf.py
+++ b/src/transformers/models/llama/convert_llama_weights_to_hf.py
@@ -59,17 +59,22 @@ INTERMEDIATE_SIZE_MAP = {
"13B": 13824,
"30B": 17920,
"65B": 22016,
+ "70B": 28672,
}
NUM_SHARDS = {
"7B": 1,
+ "7Bf": 1,
"13B": 2,
+ "13Bf": 2,
"30B": 4,
"65B": 8,
+ "70B": 8,
+ "70Bf": 8,
}
-def compute_intermediate_size(n):
- return int(math.ceil(n * 8 / 3) + 255) // 256 * 256
+def compute_intermediate_size(n, ffn_dim_multiplier=1):
+ return int((math.ceil(n * 8 / 3) + 255) * ffn_dim_multiplier // 256 * 256)
def read_json(path):
@@ -82,7 +87,7 @@ def write_json(text, path):
json.dump(text, f)
-def write_model(model_path, input_base_path, model_size):
+def write_model(model_path, input_base_path, model_size, safe_serialization=True):
os.makedirs(model_path, exist_ok=True)
tmp_model_path = os.path.join(model_path, "tmp")
os.makedirs(tmp_model_path, exist_ok=True)
@@ -97,9 +102,17 @@ def write_model(model_path, input_base_path, model_size):
base = 10000.0
inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
+ if "n_kv_heads" in params:
+ num_key_value_heads = params["n_kv_heads"] # for GQA / MQA
+ num_local_key_value_heads = n_heads_per_shard // num_key_value_heads
+ key_value_dim = dim//num_key_value_heads
+ else: # compatibility with other checkpoints
+ num_key_value_heads = n_heads
+ num_local_key_value_heads = n_heads_per_shard
+ key_value_dim = dim
# permute for sliced rotary
- def permute(w):
- return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim)
+ def permute(w, n_heads = n_heads,dim1=dim, dim2=dim):
+ return w.view(n_heads, dim1 // n_heads // 2, 2, dim2).transpose(1, 2).reshape(dim1, dim2)
print(f"Fetching all parameters from the checkpoint at {input_base_path}.")
# Load weights
@@ -160,19 +173,19 @@ def write_model(model_path, input_base_path, model_size):
state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
torch.cat(
[
- loaded[i][f"layers.{layer_i}.attention.wk.weight"].view(n_heads_per_shard, dims_per_head, dim)
+ loaded[i][f"layers.{layer_i}.attention.wk.weight"].view(num_local_key_value_heads, dims_per_head, dim)
for i in range(num_shards)
],
dim=0,
- ).reshape(dim, dim)
+ ).reshape(key_value_dim, dim),num_key_value_heads, key_value_dim, dim
)
state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = torch.cat(
[
- loaded[i][f"layers.{layer_i}.attention.wv.weight"].view(n_heads_per_shard, dims_per_head, dim)
+ loaded[i][f"layers.{layer_i}.attention.wv.weight"].view(num_local_key_value_heads, dims_per_head, dim)
for i in range(num_shards)
],
dim=0,
- ).reshape(dim, dim)
+ ).reshape(key_value_dim, dim)
state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = torch.cat(
[loaded[i][f"layers.{layer_i}.attention.wo.weight"] for i in range(num_shards)], dim=1
@@ -218,13 +231,14 @@ def write_model(model_path, input_base_path, model_size):
# Write configs
index_dict["metadata"] = {"total_size": param_count * 2}
write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))
-
+ ffn_dim_multiplier = params["ffn_dim_multiplier"] if "ffn_dim_multiplier" in params else 1
config = LlamaConfig(
hidden_size=dim,
- intermediate_size=compute_intermediate_size(dim),
+ intermediate_size=compute_intermediate_size(dim, ffn_dim_multiplier),
num_attention_heads=params["n_heads"],
num_hidden_layers=params["n_layers"],
rms_norm_eps=params["norm_eps"],
+ num_key_value_heads = num_key_value_heads
)
config.save_pretrained(tmp_model_path)
@@ -239,7 +253,7 @@ def write_model(model_path, input_base_path, model_size):
del model.config._name_or_path
print("Saving in the Transformers format.")
- model.save_pretrained(model_path)
+ model.save_pretrained(model_path, safe_serialization=safe_serialization)
shutil.rmtree(tmp_model_path)
@@ -259,18 +273,20 @@ def main():
)
parser.add_argument(
"--model_size",
- choices=["7B", "13B", "30B", "65B", "tokenizer_only"],
+ choices=["7B", "7Bf","13B", "13Bf", "30B", "65B", "70B", "70Bf", "tokenizer_only"],
)
parser.add_argument(
"--output_dir",
help="Location to write HF model and tokenizer",
)
+ parser.add_argument("--safe_serialization",type=bool, help="Whether or not to save using `safetensors`.")
args = parser.parse_args()
if args.model_size != "tokenizer_only":
write_model(
model_path=args.output_dir,
input_base_path=os.path.join(args.input_dir, args.model_size),
model_size=args.model_size,
+ safe_serialization=args.safe_serialization
)
spm_path = os.path.join(args.input_dir, "tokenizer.model")
write_tokenizer(args.output_dir, spm_path)
diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py
index 6cdbb2623..d70d0e00d 100755
--- a/src/transformers/models/llama/modeling_llama.py
+++ b/src/transformers/models/llama/modeling_llama.py
@@ -85,7 +85,7 @@ class LlamaRMSNorm(nn.Module):
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
- return (self.weight * hidden_states).to(input_dtype)
+ return self.weight.to(input_dtype) * hidden_states
class LlamaRotaryEmbedding(torch.nn.Module):
@@ -204,6 +204,16 @@ class LlamaMLP(nn.Module):
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+ """torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
+ bs, n_kv_heads, slen, head_dim = hidden_states.shape
+ if n_rep == 1:
+ return hidden_states
+ hidden_states = hidden_states[:, :, None, :, :].expand(bs, n_kv_heads, n_rep, slen, head_dim)
+ return hidden_states.reshape(bs, n_kv_heads * n_rep, slen, head_dim)
+
+
+
class LlamaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
@@ -213,6 +223,8 @@ class LlamaAttention(nn.Module):
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
+ self.num_key_value_heads = config.num_key_value_heads
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
if (self.head_dim * self.num_heads) != self.hidden_size:
@@ -221,8 +233,8 @@ class LlamaAttention(nn.Module):
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
- self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
- self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
self._init_rope()
@@ -243,9 +255,6 @@ class LlamaAttention(nn.Module):
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
- def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
- return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
-
def forward(
self,
hidden_states: torch.Tensor,
@@ -258,8 +267,8 @@ class LlamaAttention(nn.Module):
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
- key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
- value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+ key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+ value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
kv_seq_len = key_states.shape[-2]
if past_key_value is not None:
@@ -275,6 +284,9 @@ class LlamaAttention(nn.Module):
past_key_value = (key_states, value_states) if use_cache else None
+ # repeat k/v heads if n_kv_heads < n_heads
+ key_states = repeat_kv(key_states, self.num_key_value_groups) # (bs, n_heads, seqlen, head_dim)
+ value_states = repeat_kv(value_states, self.num_key_value_groups) # (bs, n_heads, seqlen, head_dim)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
diff --git a/src/transformers/models/llama/tokenization_llama.py b/src/transformers/models/llama/tokenization_llama.py
index 193d4edd5..f0fa81c3e 100644
--- a/src/transformers/models/llama/tokenization_llama.py
+++ b/src/transformers/models/llama/tokenization_llama.py
@@ -21,13 +21,15 @@
"""Tokenization classes for LLaMA."""
import os
from shutil import copyfile
-from typing import Any, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
import sentencepiece as spm
from ...tokenization_utils import AddedToken, PreTrainedTokenizer
from ...utils import logging
+if TYPE_CHECKING:
+ from transformers.pipelines.conversational import Conversation
logger = logging.get_logger(__name__)
@@ -46,6 +48,7 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
}
SPIECE_UNDERLINE = "▁"
+B_INST, E_INST = "[INST]", "[/INST]"
class LlamaTokenizer(PreTrainedTokenizer):
"""
@@ -314,3 +317,34 @@ class LlamaTokenizer(PreTrainedTokenizer):
output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
return output
+
+ def _build_conversation_input_ids(self, conversation: "Conversation") -> List[int]:
+ """Builds the input ids for a conversation.
+ This is the format used in the provided examples. "
+ ```
+ <bos>[INST] Prompt [/INST] Answer <eos>
+ <bos>[INST] Prompt [/INST]
+ ```
+ Args:
+ conversation (`Conversation`):
+ Conversation to build input ids for.
+ Returns:
+ `List[int]`:
+ Input ids for the conversation.
+ """
+ dialogue = list(conversation.iter_texts())
+ if not all([is_user for is_user, msg in dialogue[::2]]) or not all([not is_user for is_user, msg in dialogue[1::2]]):
+ raise ValueError(
+ "The model only supports 'user' and 'assistant' roles, starting with user and alternating (u/a/u/a/u...)"
+ )
+ dialog_tokens: List[int] = sum(
+ [
+ [self.bos_token_id]+self.encode(f"{B_INST} {(prompt[1]).strip()} {E_INST} {(answer[1]).strip()} ", add_special_tokens = False) + [self.eos_token_id]
+ for prompt, answer in zip(dialogue[::2], dialogue[1::2])
+ ],
+ [],
+ )
+ if not (dialogue[-1][0]):
+ raise ValueError(f"Last message must be from user, got {dialogue[-1]['role']}")
+ dialog_tokens += [self.bos_token_id] + self.encode(f"{B_INST} {(dialogue[-1][1]).strip()} {E_INST}", add_special_tokens = False)
+ return dialog_tokens
\ No newline at end of file
diff --git a/src/transformers/models/llama/tokenization_llama_fast.py b/src/transformers/models/llama/tokenization_llama_fast.py
index 28e9413a5..5a3127c69 100644
--- a/src/transformers/models/llama/tokenization_llama_fast.py
+++ b/src/transformers/models/llama/tokenization_llama_fast.py
@@ -33,6 +33,12 @@ else:
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model", "tokenizer_file": "tokenizer.json"}
+B_INST, E_INST = "[INST]", "[/INST]"
+B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
+DEFAULT_SYSTEM_PROMPT = """\
+You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."""
class LlamaTokenizerFast(PreTrainedTokenizerFast):
"""
@@ -171,3 +177,43 @@ class LlamaTokenizerFast(PreTrainedTokenizerFast):
copyfile(self.vocab_file, out_vocab_file)
return (out_vocab_file,)
+
+ def _build_conversation_input_ids(self, conversation: "Conversation") -> List[int]:
+ """Builds the input ids for a conversation.
+ This is the format used in the provided examples. System prompts should be manually added
+ at the beginning of the conversation. If no system prompt is given, the `DEFAULT_SYSTEM_PROMPT` will
+ be used.
+ ```
+ <bos>[INST] Prompt [/INST] Answer <eos>
+ <bos>[INST] Prompt [/INST]
+ ```
+ Args:
+ conversation (`Conversation`):
+ Conversation to build input ids for.
+ Returns:
+ `List[int]`:
+ Input ids for the conversation.
+ """
+ dialogue = list(conversation.iter_texts())
+ if not all([is_user for is_user, msg in dialogue[::2]]) or not all([not is_user for is_user, msg in dialogue[1::2]]):
+ raise ValueError(
+ "The model only supports 'user' and 'assistant' roles, starting with user and alternating (u/a/u/a/u...)"
+ )
+
+ # TODO add system prompt
+ dialog_tokens: List[int] = []
+ if B_SYS not in conversation.past_user_inputs[0]:
+ conversation.past_user_inputs[0] = B_SYS + DEFAULT_SYSTEM_PROMPT + E_SYS + conversation.past_user_inputs[0]
+
+
+ dialog_tokens += sum(
+ [
+ [self.bos_token_id]+self.encode(f"{B_INST} {(prompt[1]).strip()} {E_INST} {(answer[1]).strip()} ", add_special_tokens = False) + [self.eos_token_id]
+ for prompt, answer in zip(dialogue[::2], dialogue[1::2])
+ ],
+ [],
+ )
+ if not (dialogue[-1][0]):
+ raise ValueError(f"Last message must be from user, got {dialogue[-1]['role']}")
+ dialog_tokens += [self.bos_token_id] + self.encode(f"{B_INST} {(dialogue[-1][1]).strip()} {E_INST}", add_special_tokens = False)
+ return dialog_tokens
\ No newline at end of file
diff --git a/tests/models/llama/test_modeling_llama.py b/tests/models/llama/test_modeling_llama.py
index e8b808461..a43ba3654 100644
--- a/tests/models/llama/test_modeling_llama.py
+++ b/tests/models/llama/test_modeling_llama.py
@@ -365,3 +365,65 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
# The output should be different for long inputs
self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
+class LlamaIntegrationTest(unittest.TestCase):
+
+ def test_model_7b_logits(self):
+ input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+ model = LlamaForCausalLM.from_pretrained("/raid/arthur/llama-7b", device_map = "auto")
+ out = model(torch.tensor(input_ids))
+ # Expected mean on dim = -1
+ EXPECTED_MEAN = torch.tensor([[-6.6550, -4.1227, -4.9859, -3.2406, 0.8262, -3.0033, 1.2964, -3.3699]])
+ # slicing logits[0, 0, 0:30]
+ EXPECTED_SLICE = torch.tensor([-12.8281, -7.4453, -0.4639, -8.0625, -7.2500, -8.0000, -6.4883,
+ -7.7695, -7.8438, -7.0312, -6.2188, -7.1328, -1.8496, 1.9961,
+ -8.6250, -6.7227, -12.8281, -6.9492, -7.0742, -7.7852, -7.5820,
+ -7.9062, -6.9375, -7.9805, -8.3438, -8.1562, -8.0469, -7.6250,
+ -7.7422, -7.3398])
+
+ def test_model_7bf_logits(self):
+ input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+ model = LlamaForCausalLM.from_pretrained("/raid/arthur/llama-7bf", device_map = "auto")
+ out = model(torch.tensor(input_ids))
+ # Expected mean on dim = -1
+ EXPECTED_MEAN = torch.tensor([ 0.0719, -4.1667, -3.4864, -4.6226, 1.7280, -3.6511, 1.0122, -0.1268])
+ # slicing logits[0, 0, 0:30]
+ EXPECTED_SLICE = torch.tensor([ 0.1038, -0.2218, 0.3132, -0.8379, 1.5576, 2.6680, 1.5811, 2.5078,
+ 1.2129, 0.3484, 1.6602, 0.8213, 0.6294, 0.4907, 1.2588, 0.3982,
+ 0.1039, 1.9062, 0.6665, 1.0439, 0.5850, 1.8535, 2.3828, 1.8096,
+ 1.0498, 1.4629, 1.3506, 2.8574, 1.3447, 1.9971])
+
+ def test_model_13b_logits(self):
+ input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+ model = LlamaForCausalLM.from_pretrained("/raid/arthur/llama-13b", device_map = "auto")
+ out = model(torch.tensor(input_ids))
+ # Expected mean on dim = -1
+ EXPECTED_MEAN = torch.tensor([[-2.0622, -1.2794, -1.1638, -0.9788, -1.4603, -1.0238, -1.7893, -1.4411]],dtype=torch.float32)
+ # slicing logits[0, 0, 0:30]
+ EXPECTED_SLICE = torch.tensor([-8.1406, -8.0547, 2.7461, -1.2344, -0.1448, -1.8262, -1.0020, -1.8154,
+ -1.6895, -1.8516, -2.3574, -0.9277, 3.7598, 6.5742, -1.2998, -0.1177,
+ -8.1406, -2.9688, -2.9199, -3.1699, -3.5254, -2.3555, -2.7988, -3.4141,
+ -2.8262, -4.5195, -3.3379, -3.3164, -2.7832, -3.0273])
+
+
+
+ def test_model_13bf_logits(self):
+ input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+ model = LlamaForCausalLM.from_pretrained("/raid/arthur/llama-13bf", device_map = "auto")
+ out = model(torch.tensor(input_ids))
+ # Expected mean on dim = -1
+ EXPECTED_MEAN = torch.tensor([[-0.8562, -1.8520, -0.7551, -0.4162, -1.5161, -1.2038, -2.4823, -2.3254]])
+ # slicing logits[0, 0, 0:30]
+ EXPECTED_SLICE = torch.tensor([-2.2227, 4.8828, 0.9023, -0.4578, -0.7871, -0.1033, -0.6221, -0.5786,
+ -0.7803, -1.0674, -1.2920, -0.1570, 0.8008, 2.0723, -0.9497, 0.2771,
+ -2.2227, -0.7612, -1.4346, -1.2061, -1.6426, -0.3000, -0.7139, -1.1934,
+ -1.8691, -1.6973, -1.5947, -1.2705, -0.3523, -0.5513])
+
+ def test_model_70b_logits(self):
+ input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
+
+ EXPECTED_MEAN = torch.tensor([-9.4922, -3.9551, 1.7998, -5.6758, -5.1055, -5.8984, -4.8320, -6.8086,
+ -6.5391, -5.6172, -5.5820, -5.5352, 1.7881, 3.6289, -6.5117, -3.4785,
+ -9.5000, -6.0352, -6.8125, -6.0195, -6.6836, -5.4727, -6.2812, -6.0391,
+ -7.3398, -7.4297, -7.4844, -6.5820, -5.8789, -5.5312],dtype=torch.float32)
+ EXPECTED_SLICE = torch.tensor([[-4.2327, -3.3360, -4.6665, -4.7631, -1.8180, -3.4170, -1.4211, -3.1810]],dtype=torch.float32)
+