scaffold

Author: lucidrains

.github/workflows/python-publish.yml

@@ -0,0 +1,38 @@
+
+  
+# This workflow will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.x'
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install build
+    - name: Build package
+      run: python -m build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@27b31702a0e7fc50959f5ad993c78deac1bdfc29
+      with:
+        user: __token__
+        password: ${{ secrets.PYPI_API_TOKEN }}

README.md

@@ -1,2 +1,23 @@
-# PaLM-rlhf-pytorch
+## PaLM + RLHF - Pytorch (wip)
+
 Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture
+
+## Citations
+
+```bibtex
+@article{Stiennon2020LearningTS,
+    title   = {Learning to summarize from human feedback},
+    author  = {Nisan Stiennon and Long Ouyang and Jeff Wu and Daniel M. Ziegler and Ryan J. Lowe and Chelsea Voss and Alec Radford and Dario Amodei and Paul Christiano},
+    journal = {ArXiv},
+    year    = {2020},
+    volume  = {abs/2009.01325}
+}
+```
+
+```bibtex
+@inproceedings{Chowdhery2022PaLMSL,
+    title   = {PaLM: Scaling Language Modeling with Pathways},
+    author  = {Aakanksha Chowdhery and Sharan Narang and Jacob Devlin and Maarten Bosma and Gaurav Mishra and Adam Roberts and Paul Barham and Hyung Won Chung and Charles Sutton and Sebastian Gehrmann and Parker Schuh and Kensen Shi and Sasha Tsvyashchenko and Joshua Maynez and Abhishek Rao and Parker Barnes and Yi Tay and Noam M. Shazeer and Vinodkumar Prabhakaran and Emily Reif and Nan Du and Benton C. Hutchinson and Reiner Pope and James Bradbury and Jacob Austin and Michael Isard and Guy Gur-Ari and Pengcheng Yin and Toju Duke and Anselm Levskaya and Sanjay Ghemawat and Sunipa Dev and Henryk Michalewski and Xavier Garc{\'i}a and Vedant Misra and Kevin Robinson and Liam Fedus and Denny Zhou and Daphne Ippolito and David Luan and Hyeontaek Lim and Barret Zoph and Alexander Spiridonov and Ryan Sepassi and David Dohan and Shivani Agrawal and Mark Omernick and Andrew M. Dai and Thanumalayan Sankaranarayana Pillai and Marie Pellat and Aitor Lewkowycz and Erica Oliveira Moreira and Rewon Child and Oleksandr Polozov and Katherine Lee and Zongwei Zhou and Xuezhi Wang and Brennan Saeta and Mark Diaz and Orhan Firat and Michele Catasta and Jason Wei and Kathleen S. Meier-Hellstern and Douglas Eck and Jeff Dean and Slav Petrov and Noah Fiedel},
+    year    = {2022}
+}
+```

palm_rlhf_pytorch/__init__.py

@@ -0,0 +1,197 @@
+import torch
+from torch import einsum, nn
+import torch.nn.functional as F
+
+from einops import rearrange
+
+# normalization
+# they use layernorm without bias, something that pytorch does not offer
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.gamma = nn.Parameter(torch.ones(dim))
+        self.register_buffer("beta", torch.zeros(dim))
+
+    def forward(self, x):
+        return F.layer_norm(x, x.shape[-1:], self.gamma, self.beta)
+
+# residual
+
+
+class Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x):
+        return self.fn(x) + x
+
+
+# rotary positional embedding
+# https://arxiv.org/abs/2104.09864
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+    def forward(self, max_seq_len, *, device):
+        seq = torch.arange(max_seq_len, device=device, dtype=self.inv_freq.dtype)
+        freqs = einsum("i , j -> i j", seq, self.inv_freq)
+        return torch.cat((freqs, freqs), dim=-1)
+
+
+def rotate_half(x):
+    x = rearrange(x, "... (j d) -> ... j d", j=2)
+    x1, x2 = x.unbind(dim=-2)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(pos, t):
+    return (t * pos.cos()) + (rotate_half(t) * pos.sin())
+
+
+def l2norm(t):
+    return F.normalize(t, dim = -1)
+
+# classic Noam Shazeer paper, except here they use SwiGLU instead of the more popular GEGLU for gating the feedforward
+# https://arxiv.org/abs/2002.05202
+
+
+class SwiGLU(nn.Module):
+    def forward(self, x):
+        x, gate = x.chunk(2, dim=-1)
+        return F.silu(gate) * x
+
+
+# parallel attention and feedforward with residual
+# discovered by Wang et al + EleutherAI from GPT-J fame
+
+
+class ParallelTransformerBlock(nn.Module):
+    def __init__(self, dim, dim_head=64, heads=8, ff_mult=4):
+        super().__init__()
+        self.norm = LayerNorm(dim)
+
+        attn_inner_dim = dim_head * heads
+        ff_inner_dim = dim * ff_mult
+        self.fused_dims = (attn_inner_dim, dim_head, dim_head, (ff_inner_dim * 2))
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+        self.rotary_emb = RotaryEmbedding(dim_head)
+
+        self.fused_attn_ff_proj = nn.Linear(dim, sum(self.fused_dims), bias=False)
+        self.attn_out = nn.Linear(attn_inner_dim, dim, bias=False)
+
+        self.ff_out = nn.Sequential(
+            SwiGLU(),
+            nn.Linear(ff_inner_dim, dim, bias=False)
+        )
+
+        # for caching causal mask and rotary embeddings
+
+        self.register_buffer("mask", None, persistent=False)
+        self.register_buffer("pos_emb", None, persistent=False)
+
+    def get_mask(self, n, device):
+        if self.mask is not None and self.mask.shape[-1] >= n:
+            return self.mask[:n, :n]
+
+        mask = torch.ones((n, n), device=device, dtype=torch.bool).triu(1)
+        self.register_buffer("mask", mask, persistent=False)
+        return mask
+
+    def get_rotary_embedding(self, n, device):
+        if self.pos_emb is not None and self.pos_emb.shape[-2] >= n:
+            return self.pos_emb[:n]
+
+        pos_emb = self.rotary_emb(n, device=device)
+        self.register_buffer("pos_emb", pos_emb, persistent=False)
+        return pos_emb
+
+    def forward(self, x):
+        """
+        einstein notation
+        b - batch
+        h - heads
+        n, i, j - sequence length (base sequence length, source, target)
+        d - feature dimension
+        """
+
+        n, device, h = x.shape[1], x.device, self.heads
+
+        # pre layernorm
+
+        x = self.norm(x)
+
+        # attention queries, keys, values, and feedforward inner
+
+        q, kv, v, ff = self.fused_attn_ff_proj(x).split(self.fused_dims, dim=-1)
+
+
+        # split heads
+        # they use multi-query single-key-value attention, yet another Noam Shazeer paper
+        # they found no performance loss past a certain scale, and more efficient decoding obviously
+        # https://arxiv.org/abs/1911.02150
+
+        q = rearrange(q, "b n (h d) -> b h n d", h=h)
+
+        v = kv.clone()
+
+        q, kv = map(l2norm, (q, kv))
+        # rotary embeddings
+
+        positions = self.get_rotary_embedding(n, device)
+        q, k = map(lambda t: apply_rotary_pos_emb(positions, t), (q, kv))
+
+        # scale
+
+        # q = q * self.scale
+
+        # similarity
+
+        sim = einsum("b h i d, b j d -> b h i j", q, k) * 8
+
+        # causal mask
+
+        causal_mask = self.get_mask(n, device)
+        sim = sim.masked_fill(causal_mask, -torch.finfo(sim.dtype).max)
+
+        # attention
+
+        attn = sim.softmax(dim=-1)
+
+        # aggregate values
+
+        out = einsum("b h i j, b j d -> b h i d", attn, v)
+
+        # merge heads
+
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.attn_out(out) + self.ff_out(ff)
+
+
+# transformer
+
+
+def PaLM(*, dim, num_tokens, depth, dim_head=64, heads=8, ff_mult=4):
+    net = nn.Sequential(
+        nn.Embedding(num_tokens, dim),
+        *[
+            Residual(ParallelTransformerBlock(dim=dim, dim_head=dim_head, heads=heads, ff_mult=ff_mult))
+            for _ in range(depth)
+        ],
+        LayerNorm(dim),
+        nn.Linear(dim, num_tokens, bias=False)
+    )
+
+    # they used embedding weight tied projection out to logits, not common, but works
+    net[-1].weight = net[0].weight
+
+    nn.init.normal_(net[0].weight, std=0.02)
+    return net