Jax Sourceror is a Python library that allows you to recreate JAX source code from a jitted jax function (specifically its jaxpr)
and a set of inputs. This is useful for minimizing bugs, debugging, teaching, and understanding how JAX works under the hood.
The code this generates is definitely not going to be clean, idiomatic, or sometimes even correct, but it should be a good starting point for understanding what's going on.
I created it mostly as a learning exercise and to minimize bugs in framework-heavy code (i.e. removing layers of equinox or flax abstraction to get to the JAX code).
This is more of a "submit a PR" or "fork it" repo than a "this doesn't work for me" repo, but I'm happy to help out if you're stuck.
Jax Sourceror can turn this:
import jax
import jax.numpy as jnp
block_len = 64
seq_len = 128
batch = 4
num_heads = 2
embed_size = 32
num_layers = 2
head_size = 16
def pseudo_sliding_window_attention(x):
# (this is not attention, but is minimized from attn)
# dims are [batch, len, num_heads, head_dim]
# having num_heads is important. num_heads = 1, no boom
def block(block_idx):
query_block = jax.lax.dynamic_slice_in_dim(x, block_idx, block_len, axis=1)
weights = jnp.sum(query_block, axis=3) # [batch, len, num_heads]
weights = jax.lax.broadcast_in_dim(weights, (batch, block_len, num_heads, block_len),
(0, 1, 2)) # [batch, len, num_heads, len]
# weights = with_sharding_constraint(weights, P('data', None, None, None))
# without "bias", no boom
bias = jnp.ones(block_len)
bias = jnp.broadcast_to(bias, (batch, block_len, num_heads, block_len))
weights = weights + bias
return jnp.einsum('bqhk,bkhd->bqhd', weights, query_block).astype(query_block.dtype)
num_blocks = seq_len // block_len
blocked_attn = jax.lax.map(block, jnp.arange(0, num_blocks)) # [num_blocks, batch, len, num_heads, head_dim]
blocked_attn = jnp.concatenate(blocked_attn, axis=1)
return blocked_attn
def fwd(params, x):
@partial(jax.checkpoint, prevent_cse=False)
def layer(x, params):
qkv, o = params
y = jnp.einsum('bte,hde->bthd', x, qkv)
y = pseudo_sliding_window_attention(y)
z = jnp.einsum('bthd,hde->bte', y, o)
return z, None
x, _ = jax.lax.scan(layer, x, params)
return x
def loss_fn(params, x):
x = fwd(params, x)
l = jnp.mean(x)
return l
def grad_fn(params, x):
loss, grad = jax.value_and_grad(loss_fn)(params, x)
# we can't reasonably sourcerize pytrees so just get the leaves
return loss, *jax.tree_util.tree_leaves(grad)
qkv = jnp.ones((num_layers, num_heads, head_size, embed_size), dtype=jnp.bfloat16)
o = jnp.ones((num_layers, num_heads, head_size, embed_size), dtype=jnp.bfloat16)
x = jnp.ones((batch, seq_len, embed_size), dtype=jnp.bfloat16)
params = (qkv, o)
grad_fn(params, x)into this:
def grad_fn(*args, **kwargs):
def grad_fn(a, b, c):
d = jax.numpy.zeros((4, 128, 32), jax.numpy.bfloat16)
e = jax.numpy.ones((64,), jax.numpy.float32)
f = jax.lax.broadcast_in_dim(e, shape=(4, 64, 2, 64), broadcast_dimensions=(3,))
def fn_1(carry, x):
# (I would like to make this part nicer)
(g, h, i) = (carry, *x)
j = jax.lax.dot_general(g, h, (((2,), (2,)), ((), ())), None, jax.numpy.bfloat16)
def fn_2(k, l):
def fn_3(carry, x):
(m,) = (*carry, x)
n = jax.lax.dynamic_slice(l, (0, m, 0, 0), slice_sizes=(4, 64, 2, 16))
o = n.astype(jax.numpy.float32)
p = jax.numpy.sum(o, axis=(3,))
q = p.astype(jax.numpy.bfloat16)
r = jax.lax.broadcast_in_dim(q, shape=(4, 64, 2, 64), broadcast_dimensions=(0, 1, 2))
s = r.astype(jax.numpy.float32)
t = s + k
u = jax.lax.dot_general(n, t, (((1,), (3,)), ((0, 2), (0, 2))), None, jax.numpy.float32)
v = jax.lax.transpose(u, permutation=(0, 3, 1, 2))
w = v.astype(jax.numpy.bfloat16)
return ((), w)
(final_carry, ys) = jax.lax.scan(fn_3, (), jax.numpy.array([0, 1], dtype=jax.numpy.int32), length=2, unroll=1, reverse=False)
x = ys
return x
y = fn_2(f, j)
z = jax.numpy.reshape(jax.numpy.transpose(y, (1, 0, 2, 3, 4)), (4, 128, 2, 16))
ba = jax.lax.dot_general(z, i, (((3, 2), (1, 0)), ((), ())), None, jax.numpy.bfloat16)
return (ba, g)
(final_carry, ys) = jax.lax.scan(fn_1, c, (a, b), length=2, unroll=1, reverse=False)
bb = final_carry
bc = ys
bd = bb.astype(jax.numpy.float32)
be = jax.numpy.sum(bd, axis=(0, 1, 2))
bf = be / 16384.0
bg = bf.astype(jax.numpy.bfloat16)
bh = jax.lax.broadcast_in_dim(6.103515625e-05, shape=(4, 128, 32), broadcast_dimensions=())
bi = bh.astype(jax.numpy.bfloat16)
def fn_4(carry, x):
(bj, bk, bl, bm) = (carry, *x)
def fn_5(bn, bo, bp, bq):
br = jax.lax.dot_general(bn, bo, (((2,), (2,)), ((), ())), None, jax.numpy.bfloat16)
bs = jax.numpy.ones((64,), jax.numpy.float32)
bt = jax.lax.broadcast_in_dim(bs, shape=(4, 64, 2, 64), broadcast_dimensions=(3,))
def fn_6(carry, x):
(bu,) = (*carry, x)
bv = bu < 0
bw = bu + 128
bx = jax.lax.select_n(bv, bu, bw)
by = jax.lax.dynamic_slice(br, (0, bx, 0, 0), slice_sizes=(4, 64, 2, 16))
bz = by.astype(jax.numpy.float32)
ca = jax.numpy.sum(bz, axis=(3,))
cb = ca.astype(jax.numpy.bfloat16)
cc = jax.lax.broadcast_in_dim(cb, shape=(4, 64, 2, 64), broadcast_dimensions=(0, 1, 2))
cd = cc.astype(jax.numpy.float32)
ce = cd + bt
cf = jax.lax.dot_general(by, ce, (((1,), (3,)), ((0, 2), (0, 2))), None, jax.numpy.float32)
cg = jax.lax.transpose(cf, permutation=(0, 3, 1, 2))
ch = cg.astype(jax.numpy.bfloat16)
return ((), (ch, bx, ce, by))
(final_carry, ys) = jax.lax.scan(fn_6, (), jax.numpy.array([0, 1], dtype=jax.numpy.int32), length=2, unroll=1, reverse=False)
(ci, cj, ck, cl) = ys
cm = jax.numpy.reshape(jax.numpy.transpose(ci, (1, 0, 2, 3, 4)), (4, 128, 2, 16))
cn = jax.lax.dot_general(bq, cm, (((0, 1), (0, 1)), ((), ())), None, jax.numpy.bfloat16)
co = jax.lax.transpose(cn, permutation=(1, 2, 0))
cp = jax.lax.dot_general(bq, bp, (((2,), (2,)), ((), ())), None, jax.numpy.bfloat16)
cq = jax.numpy.reshape(cp, (4, 2, 64, 2, 16))
cr = jax.lax.transpose(cq, permutation=(1, 0, 2, 3, 4))
cs = jax.numpy.zeros((4, 128, 2, 16), jax.numpy.bfloat16)
def fn_7(carry, x):
(ct, cu, cv, cw, cx) = (carry, *x)
cy = cu.astype(jax.numpy.float32)
cz = jax.lax.transpose(cy, permutation=(0, 2, 3, 1))
da = jax.lax.dot_general(cz, cx, (((2,), (3,)), ((0, 1), (0, 2))), None, jax.numpy.float32)
db = jax.lax.transpose(da, permutation=(0, 2, 1, 3))
dc = db.astype(jax.numpy.bfloat16)
dd = jax.numpy.sum(dc, axis=(3,))
de = dd.astype(jax.numpy.float32)
df = jax.lax.broadcast_in_dim(de, shape=(4, 64, 2, 16), broadcast_dimensions=(0, 1, 2))
dg = df.astype(jax.numpy.bfloat16)
dh = jax.lax.dot_general(cz, cw, (((3,), (1,)), ((0, 1), (0, 2))), None, jax.numpy.float32)
di = jax.lax.transpose(dh, permutation=(0, 3, 1, 2))
dj = di.astype(jax.numpy.bfloat16)
dk = dg + dj
dl = jax.numpy.zeros((4, 128, 2, 16), jax.numpy.bfloat16)
dm = jax.lax.dynamic_update_slice(dl, dk, (0, cv, 0, 0))
dn = ct + dm
return (dn, ())
(final_carry, ys) = jax.lax.scan(fn_7, cs, (cr, cj, ck, cl), length=2, unroll=1, reverse=True)
do = final_carry
dp = jax.lax.dot_general(do, bn, (((0, 1), (0, 1)), ((), ())), None, jax.numpy.bfloat16)
dq = jax.lax.dot_general(do, bo, (((2, 3), (0, 1)), ((), ())), None, jax.numpy.bfloat16)
return (dq, dp, co)
ckpt_fn_5 = jax.checkpoint(fn_5)
(dr, ds, dt) = ckpt_fn_5(bk, bl, bm, bj)
return (dr, (ds, dt))
(final_carry, ys) = jax.lax.scan(fn_4, bi, (bc, a, b), length=2, unroll=1, reverse=True)
du = final_carry
(dv, dw) = ys
return (bg, dv, dw)
return grad_fn(*jax.tree_leaves((args, kwargs)))Is this pretty code? No. Is it even readable? If you try hard enough. Is it correct? I think so. (It definitely passes my unit test!)
from jax_sourceror import sourcerize
source_code = sourcerize(grad_fn, *args, **kwargs)
print(source_code)