jburtoft's picture
v1.1.0: fused attention block for L = 768 / 1024
fb58c6c
Raw History Blame Contribute Delete
10.7 kB
"""PyTorch Native wrappers and model patching for NewBERT encoders (torch.compile(backend="neuron")).
Inference only. LNC is read once, at import, from NEURON_LOGICAL_NC_CONFIG (default 2, the trn2
default); set it before importing this package. Shapes the kernels do not support fall back to the
stock modules at trace time (shapes are static under torch.compile), so patching is always safe.
"""
import math
import os
import types
import torch
from torch_neuronx import nki_op
from .attn_block_kernel import attn_block_kernel
from .attn_block_pack import pack as _pack_attn
from .gated_mlp_kernel import gated_mlp_ln_kernel, gated_mlp_ln_stream_kernel, mlp_weights_fit
from .res_ln_kernel import res_ln_kernel
LNC = int(os.environ.get("NEURON_LOGICAL_NC_CONFIG", "2"))
# Fused attention block: supported sequence lengths (multiples of 128, <= 1024).
ATTN_LENGTHS = (128, 256, 384, 512, 768, 1024)
# Lengths not in ATTN_LENGTHS from this length up use nkilib attention_cte (key padding as sequence bounds)
# for the attention core; below it the compiler's own attention is faster.
CTE_MIN_LEN = 768
# Key-bias clamp: padded keys get this additive bias (exp underflows to 0 in fp32 and bf16).
MASK_BIAS = -30000.0
# ---------------------------------------------------------------- ops
@nki_op("gte_enc::res_ln", mutates_args={})
def res_ln(x: torch.Tensor, r: torch.Tensor, g: torch.Tensor, b: torch.Tensor, eps: float) -> torch.Tensor:
"""LayerNorm(x + r) * g + b. x, r [R, H]; g, b [128, H] (pre-broadcast). R % (128 * LNC) == 0."""
return res_ln_kernel[LNC](x, r, g, b, eps, True)
@nki_op("gte_enc::ln", mutates_args={})
def ln(x: torch.Tensor, g: torch.Tensor, b: torch.Tensor, eps: float) -> torch.Tensor:
"""LayerNorm(x) * g + b (no residual). Same constraints as res_ln."""
return res_ln_kernel[LNC](x, x, g, b, eps, False)
@nki_op("gte_enc::gated_mlp_ln", mutates_args={})
def gated_mlp_ln(x: torch.Tensor, wupT: torch.Tensor, wgT: torch.Tensor, wdT: torch.Tensor,
bd: torch.Tensor, g: torch.Tensor, b: torch.Tensor, eps: float) -> torch.Tensor:
"""LayerNorm(x + (gelu_tanh(x Wg^T) * x Wu^T) Wd^T + bd) * g + b. R % (128 * LNC) == 0.
Weights stay resident in SBUF when they fit (hidden 768 / intermediate 3072), else they are streamed."""
K, I = wupT.shape
if mlp_weights_fit(K, I, wupT.element_size()):
return gated_mlp_ln_kernel[LNC](x, wupT, wgT, wdT, bd, g, b, eps)
return gated_mlp_ln_stream_kernel[LNC](x, wupT, wgT, wdT, bd, g, b, eps)
@nki_op("gte_enc::attn_block", mutates_args={})
def attn_block(x: torch.Tensor, wqk: torch.Tensor, bqk: torch.Tensor, wv: torch.Tensor, bv: torch.Tensor,
wo: torch.Tensor, bo: torch.Tensor, g: torch.Tensor, b: torch.Tensor, cs: torch.Tensor,
kb: torch.Tensor, eps: float) -> torch.Tensor:
"""attn_ln(x + o_proj(attention(rope(q), rope(k), v))) for packed weights (attn_block_pack.pack)."""
return attn_block_kernel[LNC](x, wqk, bqk, wv, bv, wo, bo, g, b, cs, kb, eps)
def _bcast(v, H):
return v.detach().float().reshape(1, H).expand(128, H).contiguous()
# ---------------------------------------------------------------- modules
class NkiLN(torch.nn.Module):
"""Drop-in for nn.LayerNorm(H) over the last dim, with an optional fused residual."""
def __init__(self, ln_mod: torch.nn.LayerNorm):
super().__init__()
H = ln_mod.normalized_shape[0]
self.H, self.eps = H, float(ln_mod.eps)
self.register_buffer("g", _bcast(ln_mod.weight, H))
self.register_buffer("b", _bcast(ln_mod.bias, H))
def forward(self, x, residual=None):
shp = x.shape
x2 = x.reshape(-1, self.H)
if x2.shape[0] % (128 * LNC) != 0:
s = x if residual is None else x + residual
return torch.nn.functional.layer_norm(s, (self.H,), self.g[0].to(s.dtype),
self.b[0].to(s.dtype), self.eps)
if residual is None:
y = ln(x2, self.g, self.b, self.eps)
else:
y = res_ln(x2, residual.reshape(-1, self.H), self.g, self.b, self.eps)
return y.reshape(shp)
class NkiMlpLN(torch.nn.Module):
"""Replaces NewGatedMLP + mlp_ln: y = LN(x + down(gelu_tanh(gate(x)) * up(x)))."""
def __init__(self, mlp, ln_mod):
super().__init__()
I = mlp.intermediate_size
w = mlp.up_gate_proj.weight.detach() # [2I, H]: rows [0:I] up, [I:2I] gate
self.register_buffer("wupT", w[:I].t().contiguous())
self.register_buffer("wgT", w[I:].t().contiguous())
self.register_buffer("wdT", mlp.down_proj.weight.detach().t().contiguous())
H = self.wupT.shape[0]
self.register_buffer("bd", _bcast(mlp.down_proj.bias, H))
self.register_buffer("g", _bcast(ln_mod.weight, H))
self.register_buffer("b", _bcast(ln_mod.bias, H))
self.eps = float(ln_mod.eps)
self.H = H
self._mlp, self._ln = [mlp], [ln_mod] # stock fallback (not registered as children)
def forward(self, x):
shp = x.shape
x2 = x.reshape(-1, self.H)
if x2.shape[0] % (128 * LNC) != 0:
return self._ln[0](x + self._mlp[0](x))
return gated_mlp_ln(x2, self.wupT, self.wgT, self.wdT, self.bd, self.g, self.b, self.eps).reshape(shp)
class NkiAttnBlock(torch.nn.Module):
"""attn_ln(x + attention(x)) for one NewLayer; RoPE tables packed per supported length."""
def __init__(self, attn, ln_mod, rotary, lengths=ATTN_LENGTHS):
super().__init__()
H = attn.hidden_size
self.H, self.eps = H, float(ln_mod.eps)
sc = 1.0 / math.sqrt(attn.attention_head_size)
f = lambda t: t.detach().float().cpu().numpy()
cos, sin = rotary(torch.zeros(1, dtype=torch.float32), seq_len=max(lengths))
pk = _pack_attn(f(attn.qkv_proj.weight), f(attn.qkv_proj.bias), f(attn.o_proj.weight),
f(attn.o_proj.bias), f(ln_mod.weight), f(ln_mod.bias), f(cos), f(sin), H, sc)
dt = attn.qkv_proj.weight.dtype
for k in ("wqk", "wv", "wo"):
self.register_buffer(k, torch.from_numpy(pk[k]).to(dt))
for k in ("bqk", "bv", "bo", "g", "b"):
self.register_buffer(k, torch.from_numpy(pk[k]).float())
for L in lengths:
self.register_buffer(f"cs{L}", torch.from_numpy(pk["cs"][:, :, :L].copy()).float())
self.lengths = set(lengths)
self._attn = [attn]
self.fallback_ln = NkiLN(ln_mod) # unsupported shapes: stock (or cte) attention + NKI LN
def forward(self, x, attention_bias, rope_embeds):
B, L, H = x.shape
qsplit_ok = (L // 128) % LNC == 0
if L not in self.lengths or (B % LNC and not qsplit_ok):
a = self._attn[0](x, attention_bias, rope_embeds)[0]
return self.fallback_ln(a, x)
kb = attention_bias.reshape(B, L).float().clamp(min=MASK_BIAS).to(x.dtype)
y = attn_block(x.reshape(B * L, H), self.wqk, self.bqk, self.wv, self.bv, self.wo, self.bo,
self.g, self.b, getattr(self, f"cs{L}"), kb, self.eps)
return y.reshape(B, L, H)
# ---------------------------------------------------------------- nkilib attention_cte for long sequences
_cte = None
def _cte_attention(self, q, k, v, attention_bias, head_mask):
"""q/k/v [B, L, nh, hd]; attention_bias [B, 1, 1, L] additive (0 valid, min for pad, right-padded)."""
global _cte
B, L, nh, hd = q.shape
if L < CTE_MIN_LEN or head_mask is not None or (self.training and self.dropout.p > 0):
return self._stock_attention(q, k, v, attention_bias, head_mask)
if _cte is None:
from torch_neuronx import wrap_nki
from nkilib.core.attention.attention_cte import attention_cte
_cte = wrap_nki(attention_cte)
scale = 1.0 / math.sqrt(hd)
qh = (q * scale).permute(0, 2, 1, 3).reshape(B * nh, L, hd) # bounds path needs scale=1
kh = k.permute(0, 2, 1, 3).reshape(B * nh, L, hd)
vh = v.permute(0, 2, 1, 3).reshape(B * nh, L, hd)
vl = (attention_bias.reshape(B, L) == 0).sum(-1).to(torch.float32)
bmax = vl.repeat_interleave(nh)[:, None, None].expand(B * nh, L, 1).contiguous()
bmin = torch.zeros_like(bmax)
o = _cte[LNC](qh, kh, vh, scale=1.0, causal_mask=False, tp_q=True, tp_k=True,
bound_min=bmin, bound_max=bmax)
return o.reshape(B, nh, L, hd).permute(0, 2, 1, 3).contiguous(), None
# ---------------------------------------------------------------- patching
def _layer_forward(self, hidden_states, attention_bias, rope_embeds=None, padding_inputs=None,
attention_scale=None, subset_indices=None, head_mask=None,
output_attentions=False, qkv_inputs=None):
hidden_states = self.attn_block(hidden_states, attention_bias, rope_embeds)
hidden_states = self.mlp_block(hidden_states)
return (hidden_states,)
def patch_model(model, long_attention=True, attn_lengths=ATTN_LENGTHS):
"""Patch every NewLayer / NewEmbeddings in a NewBERT (Alibaba-NLP "new-impl") model, in place:
* embedding LayerNorm -> NKI LayerNorm
* attention half of each layer -> fused attention block (L in attn_lengths, default ATTN_LENGTHS)
* gated MLP + mlp_ln -> fused gated-MLP + residual + LayerNorm
* attention core at L >= CTE_MIN_LEN -> nkilib attention_cte (if long_attention)
Call after model.to(torch.bfloat16) and before model.to("neuron"). Inference only (eval mode).
Returns the number of layers patched.
"""
assert not model.training, "patch_model is inference only; call model.eval() first"
rot = None
for mod in model.modules():
if type(mod).__name__ == "NewEmbeddings":
rot = mod.rotary_emb
mod.LayerNorm = NkiLN(mod.LayerNorm)
assert rot is not None, "no NewEmbeddings module found (is this a NewBERT model?)"
if long_attention:
for mod in model.modules():
if type(mod).__name__ == "NewAttention":
mod._stock_attention = mod._attention
mod._attention = types.MethodType(_cte_attention, mod)
n = 0
for mod in model.modules():
if type(mod).__name__ == "NewLayer":
assert "Gated" in type(mod.mlp).__name__, "only the gated (NewGatedMLP) MLP is supported"
mod.mlp_block = NkiMlpLN(mod.mlp, mod.mlp_ln)
mod.attn_block = NkiAttnBlock(mod.attention, mod.attn_ln, rot, lengths=attn_lengths)
mod.forward = types.MethodType(_layer_forward, mod)
n += 1
return n