Skip to content

qwen3: update attention to match scaling book. #86

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

Merged
jenriver merged 1 commit into from
Nov 18, 2025
Merged

qwen3: update attention to match scaling book. #86

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
36 changes: 22 additions & 14 deletions bonsai/models/qwen3/modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,14 +289,15 @@ def __init__(self, cfg: ModelConfig, *, rngs: nnx.Rngs):
self.q_norm = RMSNorm(cfg.head_dim, cfg, rngs=rngs)
self.k_norm = RMSNorm(cfg.head_dim, cfg, rngs=rngs)
self.n_rep = cfg.num_heads // cfg.num_kv_heads
self.scale = self.head_dim**-0.5
self.scale = cfg.head_dim**-0.5

@jax.named_scope("attention")
def __call__(self, x: Array, cache: LayerCache | None, segment_ids: Array) -> Array:
query_proj = shard(self.q_norm(self.q_proj(x)), self.shd_cfg.act_btnh)
key_proj = shard(self.k_norm(self.k_proj(x)), self.shd_cfg.act_btnh)
value_proj = shard(self.v_proj(x), self.shd_cfg.act_btnh)
query_proj = shard(self.q_norm(self.q_proj(x)), self.shd_cfg.act_btnh) # [B, T, N, H]
key_proj = shard(self.k_norm(self.k_proj(x)), self.shd_cfg.act_btnh) # [B, T, K, H]
value_proj = shard(self.v_proj(x), self.shd_cfg.act_btnh) # [B, T, K, H]

# RoPE and Cache Logic
left_pads = count_left_pads(segment_ids)
left_pads = shard(left_pads, P(self.shd_cfg.act_btnh[0]))
cache.start_ind.value = jnp.where(cache.start_ind.value < 0, left_pads, cache.start_ind.value)
Expand All @@ -305,25 +306,32 @@ def __call__(self, x: Array, cache: LayerCache | None, segment_ids: Array) -> Ar
query_proj = apply_rope(query_proj, sin, cos)
key_proj = apply_rope(key_proj, sin, cos)

# Update K/V cache [B, S, K, H]
slice_indices = (0, cache.cur_ind.value, 0, 0)
cache.v_cache.value = jax.lax.dynamic_update_slice(cache.v_cache.value, value_proj, slice_indices)
cache.k_cache.value = jax.lax.dynamic_update_slice(cache.k_cache.value, key_proj, slice_indices)
b, t, qh, d = query_proj.shape

# GQA
query_proj = query_proj.reshape((b, t, self.num_kv_heads, qh // self.num_kv_heads, d))
attn_logits = jnp.einsum("BTHGD,BSHD->BHGTS", query_proj, cache.k_cache.value) * self.scale
q_pos = cache.cur_ind.value + jnp.arange(t, dtype=jnp.int32)[None, :] - cache.start_ind.value[:, None]
ts = jnp.arange(cache.size, dtype=jnp.int32) # (b, cache.size)
b, t, n, h = query_proj.shape

# GQA reshape and attention logits
query_proj_gqa = query_proj.reshape((b, t, self.num_kv_heads, self.n_rep, h))
attn_logits = jnp.einsum("BTKGH,BSKH->BTSKG", query_proj_gqa, cache.k_cache.value) * self.scale

# Masking and Softmax
q_pos = cache.cur_ind.value + jnp.arange(t, dtype=jnp.int32)[None, :] - cache.start_ind.value[:, None]
ts = jnp.arange(cache.size, dtype=jnp.int32) # (cache.size,)
kv_segment_ids = (ts[None, :] >= cache.start_ind.value[:, None]) & (ts[None, :] < cache.cur_ind.value + t)
k_pos = ts[None, :] - cache.start_ind.value[:, None] # (b, cache.size)
causal_mask = k_pos[:, None, :] <= q_pos[:, :, None]
segment_mask = kv_segment_ids[:, None, :] == segment_ids[:, :, None]
final_mask = causal_mask & segment_mask # (b, q_len, k_len)
attn_logits = jnp.where(final_mask[:, None, None, :, :], attn_logits, _K_MASK)
attn_weights = jax.nn.softmax(attn_logits.astype(jnp.float32), axis=-1).astype(attn_logits.dtype)
qkv = jnp.einsum("BHGTS,BSHD->BTHGD", attn_weights, cache.v_cache.value).reshape((b, t, qh, d))
final_mask = causal_mask & segment_mask # (B, T, S)
attn_mask = final_mask[:, :, :, None, None]
attn_logits = jnp.where(attn_mask, attn_logits, _K_MASK)

# Softmax
attn_weights = jax.nn.softmax(attn_logits.astype(jnp.float32), axis=2).astype(attn_logits.dtype)
qkv = jnp.einsum("BTSKG,BSKH->BTKGH", attn_weights, cache.v_cache.value)
qkv = qkv.reshape((b, t, n, h))

cache.cur_ind.value = cache.cur_ind.value + t
return shard(self.o_proj(qkv), self.shd_cfg.act_btd)
Expand Down