facebookresearch / facebookresearch/memory

Memory implementation for 1B

Open
#4 1 comment 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
379
Forks
32
PR merge metrics
No merged PRs in 30d

Description

Hi can you please provide the config for 1B memory implementation I am trying to use content from 7B PK config but it doesnt work Thsi is the config I made dump_dir: runss/1bb_memory_model
name: "1b_pkplus_1024"
steps: 100_000
#probe_freq: 1000
seed: 777

optim:
lr: 3e-4
weight_decay: 0.01
warmup: 2000
lr_min_ratio: 0.0001
clip: 1.0
scheduler: cosine
annealing_step: 50000

distributed:
fsdp_type: full_shard
compile: true
model_dtype: bf16
matmul_allow_tf32: false
selective_activation_checkpointing: true
tp_size: 1
memory_parallel_size: 3

model:
dim: 2048
n_layers: 25
n_heads: 16
head_dim: 128
n_kv_heads: 4
rope_theta: 100_000
ffn_dim_multiplier: 1.3
multiple_of: 256
norm_eps: 1e-5
productkey_args:
is_enabled: True
layers: "6,13,20"
mem_share_values: True
mem_n_keys: 1024 # Retained, already a power of 2
value_fixed_lr: 0.0001
swilu_projection: true
mem_k_dim: 512 # Retained, already a power of 2

data:
root_dir: data/shuffled
sources:
user_dataset/arxiv: 1.0
batch_size: 24
prefetch_size: 128
seq_len: 1024
n_views: 2
load_async: true
add_bos: true
add_eos: true
tokenizer:
name: tiktoken
path: /home/gureshme/snap/firefox/memory/tokenizers/llama3/original/tokenizer.model

profiling:
run: true
mem_warmup: 100
mem_steps: 8
profile_warmup: 200
profile_steps: 8

checkpoint:
dump:
every: 5000
keep: 5

logging:
freq: 10

eval:
generator:
max_tokens: 8192
dtype: bf16
temperature: 1.0
top_p: 0.95
harness:
tasks:
- hellaswag
- piqa
- task: nq_open
num_fewshot: 5 and I keep getting this errror [rank0]: def embedding_bag_k(
[rank0]: out_ptr, # [B, dim]
[rank0]: indices_ptr, # [B, bag_size]
[rank0]: weight_ptr, # [n_keys**2, dim]
[rank0]: per_sample_weights, # [B, bag_size]
[rank0]: dim: tl.constexpr,
[rank0]: bag_size: tl.constexpr,
[rank0]: ):
[rank0]: out_idx = tl.program_id(axis=0).to(tl.int64)
[rank0]: out_value = tl.zeros([dim], dtype=tl.float32)
[rank0]: ^
[rank1]: Traceback (most recent call last):
[rank1]: File "/home/gureshme/.local/lib/python3.10/site-packages/triton/language/core.py", line 35, in wrapper
[rank1]: return fn(*args, **kwargs)
[rank1]: File "/home/gureshme/.local/lib/python3.10/site-packages/triton/language/core.py", line 1220, in full
[rank1]: shape = _shape_check_impl(shape)
[rank1]: File "/home/gureshme/.local/lib/python3.10/site-packages/triton/language/core.py", line 1205, in _shape_check_impl
[rank1]: raise ValueError(f"Shape element {i} must be a power of 2")
[rank1]: ValueError: Shape element 0 must be a power of 2

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.