Skip to content

Fix DeepSeek4HyperHead instantiation error by using mhc_reduce #4773

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
olufiyin19 wants to merge 1 commit into from
+2 −66
Closed

Fix DeepSeek4HyperHead instantiation error by using mhc_reduce #4773

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
54 changes: 0 additions & 54 deletions src/maxtext/layers/mhc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -313,58 +313,4 @@ def __call__(
return res_out + post_out, metadata


class DeepSeek4HyperHead(nnx.Module):
"""DeepSeek V4 Hyper Head for collapsing hyper-connection streams."""

def __init__(self, config: Config, rngs: nnx.Rngs):
self.config = config
self.hc_mult = config.mhc_expansion_rate
self.eps = getattr(config, "hc_eps", 1e-6)
self.dtype = config.dtype
self.weight_dtype = config.weight_dtype
self.matmul_precision = jax.lax.Precision(config.matmul_precision)

self.input_norm = RMSNorm(
num_features=self.hc_mult * config.emb_dim,
dtype=self.dtype,
weight_dtype=self.weight_dtype,
kernel_axes=("norm",),
epsilon=config.normalization_layer_epsilon,
with_scale=False,
rngs=rngs,
)

scale_init = nd_dense_init(1.0, "fan_in", "normal")
self.hc_fn = nnx.Param(
scale_init(
rngs.params(),
(self.hc_mult * config.emb_dim, self.hc_mult),
self.weight_dtype,
in_axis=0,
out_axis=1,
),
out_sharding=(None, None),
)

self.hc_base = nnx.Param(
default_bias_init(rngs.params(), (self.hc_mult,), self.weight_dtype),
out_sharding=(None,),
)
self.hc_scale = nnx.Param(
default_scalar_init(rngs.params(), (1,), self.weight_dtype),
out_sharding=(None,),
)

def __call__(self, x: Array) -> Array:
b, s, k, d = x.shape
flat = jnp.reshape(x, (b, s, k * d))
flat = self.input_norm(flat)

hc_fn = jnp.asarray(self.hc_fn[...], self.dtype)
hc_base = jnp.asarray(self.hc_base[...], self.dtype)
hc_scale = jnp.asarray(self.hc_scale[...], self.dtype)

mixes = jnp.einsum("bsm,mn -> bsn", flat, hc_fn, precision=self.matmul_precision)
pre = jax.nn.sigmoid(mixes * hc_scale + hc_base) + self.eps

return jnp.sum(jnp.expand_dims(pre, axis=-1) * x, axis=2).astype(self.dtype)
14 changes: 2 additions & 12 deletions src/maxtext/layers/nnx_decoders.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -438,13 +438,6 @@ def __init__(
self.is_gemma4 = self.config.decoder_block == DecoderBlockType.GEMMA4
self.is_gemma4_small = self.config.decoder_block == DecoderBlockType.GEMMA4_SMALL

if config.mhc_expansion_rate > 1 and config.decoder_block == DecoderBlockType.DEEPSEEK4:
self.hc_head = mhc.DeepSeek4HyperHead(
config=config,
mesh=self.mesh,
rngs=self.rngs,
)

self._init_decoder_layers(decoder_block_classes, rngs, mesh)

def _init_decoder_layers(self, decoder_block_classes, rngs, mesh):
Expand Down Expand Up @@ -1976,11 +1969,8 @@ def pure_layer_fn(graphdef_in, state_in, y_in, kv_in):

# After the final transformer layer, `y` holds the raw, un-normalized hidden state.
if getattr(cfg, "mhc_expansion_rate", 1) > 1:
if cfg.decoder_block == DecoderBlockType.DEEPSEEK4:
hidden_state = self.hc_head(y)
else:
# (batch, length, mhc_expansion_rate, emb_dim) --> (batch, length, emb_dim)
hidden_state = mhc_reduce(y)
# (batch, length, mhc_expansion_rate, emb_dim) --> (batch, length, emb_dim)
hidden_state = mhc_reduce(y)
else:
hidden_state = y

Expand Down
Loading