mirror of
https://github.com/wassname/lora-lite.git
synced 2026-06-27 16:15:50 +08:00
tidy tests to subset of metamath
This commit is contained in:
wassname
@@ -1,754 +0,0 @@
|
||||
|
||||
❯ yes make a task to deep review each one against how it's implemented. so one
|
||||
task for each. please give me the pseudocode of 1) reference code 2)
|
||||
reference paper 3) ours and compare.
|
||||
|
||||
Searched for 1 pattern, read 1 file (ctrl+o to expand)
|
||||
● 1. LoRA review
|
||||
|
||||
Paper (Hu et al. 2021):
|
||||
forward:
|
||||
h = W·x + (α/r) · B·A·x # B∈R^{d×r}, A∈R^{r×d}
|
||||
init:
|
||||
A ~ Gaussian (or kaiming), B = 0
|
||||
|
||||
peft peft_lora_layer.py:941-969:
|
||||
forward(x):
|
||||
result = base_layer(x) # W·x + bias
|
||||
for adapter in active_adapters:
|
||||
result += lora_B(lora_A(dropout(x))) * scaling # scaling = α/r or
|
||||
α/√r if rslora
|
||||
init (reset_lora_parameters, line 263):
|
||||
kaiming_uniform_(lora_A, a=√5); zeros_(lora_B)
|
||||
|
||||
ours src/lora_lite/variants/lora.py:
|
||||
param_specs:
|
||||
lora_A: (r, d_in) init=kaiming
|
||||
lora_B: (d_out, r) init=zeros
|
||||
forward(x, y): # y = base_layer(x) already computed by host
|
||||
scale = α/r
|
||||
h = einsum(x, A, "... i, r i -> ... r")
|
||||
delta = einsum(h, B, "... r, o r -> ... o")
|
||||
return y + scale·delta
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ math identical: scale·B·A·x + y
|
||||
- ✗ no dropout (peft has lora_dropout)
|
||||
- ✗ no rslora (α/√r)
|
||||
- ✗ no init_lora_weights="gaussian" option (only kaiming)
|
||||
- ✓ B=0 zero-init → identity at t=0
|
||||
|
||||
Verdict: correct vanilla LoRA. Missing dropout and rslora are deliberate scope
|
||||
cuts (not bugs). No fix needed.
|
||||
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● 2. DoRA review
|
||||
|
||||
Paper (Liu et al. 2024, §4.1, §4.3):
|
||||
W' = m · V/||V||_c, V = W + ΔV = W + (α/r)·B·A # ||·||_c = per-row L2
|
||||
init: m = ||W||_c, B = 0 → W' = W
|
||||
§4.3 trick: treat ||V+ΔV||_c as a CONSTANT (detach from grad graph)
|
||||
for stability + speed.
|
||||
|
||||
peft peft_lora_dora.py:132-166:
|
||||
forward(x):
|
||||
lora_W = B·A
|
||||
weight_norm = ||W + α/r·lora_W.detach()||_c # << DETACHED
|
||||
weight_norm = weight_norm.detach() # << belt+braces
|
||||
mag_norm_scale = (m / weight_norm)
|
||||
base_result = W·x (or passed in, minus bias)
|
||||
return (mag_norm_scale - 1)·base_result + mag_norm_scale · α/r · B·A·x
|
||||
# bias re-added outside
|
||||
|
||||
ours src/lora_lite/variants/dora.py:53-71:
|
||||
V = W + scale·B·A # << NOT detached
|
||||
v_norm = V.norm(dim=1).clamp_min(1e-12) # << NOT detached
|
||||
wx = y - bias if bias else y
|
||||
delta_lora = scale · B·A·x
|
||||
combined = wx + delta_lora
|
||||
out = (m / v_norm) · combined
|
||||
return out + bias if bias else out
|
||||
|
||||
Diff vs peft:
|
||||
- ✗ Bug: missing .detach() on the lora term inside weight_norm. Paper §4.3
|
||||
explicitly says to detach; peft detaches twice (line 144 lora_weight.detach()
|
||||
and line 152 weight_norm.detach()). Without it, grads flow through the norm
|
||||
denominator → unstable training and ~2× memory, the exact thing the paper
|
||||
warns against.
|
||||
- ✓ math otherwise equivalent: (m/||V||)·(Wx + α/r·BAx) = (m/||V||-1)·Wx +
|
||||
(m/||V||)·α/r·BAx + Wx. Both forms produce same output.
|
||||
- ✓ bias handling matches peft (subtract before scaling, add after).
|
||||
- ✓ m init ||W||_c matches.
|
||||
- = nn.Linear-only restriction is a deliberate scope cut (peft handles bnb via
|
||||
dequantize_module_weight).
|
||||
|
||||
Recommended fix: add .detach() to V (or equivalently v_norm.detach()) before
|
||||
dividing.
|
||||
|
||||
V = layer.weight + scale * BA.detach()
|
||||
v_norm = V.norm(dim=1).clamp_min(1e-12).detach()
|
||||
|
||||
Read 2 files (ctrl+o to expand)
|
||||
|
||||
● 3. PiSSA review
|
||||
|
||||
Paper (Meng et al. 2024):
|
||||
W = U S V^T (truncated to top-r)
|
||||
A = √S · V^T, B = U · √S → B·A = U S V^T
|
||||
W_res = W - B·A
|
||||
forward: W_res·x + (α/r)·B·A·x → at t=0 equals W·x
|
||||
|
||||
peft peft_lora_layer.py:360-393:
|
||||
V, S, Uh = svd(W) # V is left, Uh is right (peft
|
||||
naming)
|
||||
Sr = S[:r] / scaling # << pre-divide by α/r
|
||||
A = diag(√Sr) @ Uh[:r] # absorbs 1/√scaling
|
||||
B = V[:, :r] @ diag(√Sr) # absorbs 1/√scaling
|
||||
W_res = W - scaling · B·A # net: W - V[:,:r]·S[:r]·Uh[:r]
|
||||
(scaling cancels)
|
||||
forward: W_res·x + scaling·B·A·x
|
||||
|
||||
ours src/lora_lite/variants/pissa.py:45-76:
|
||||
U, S, Vh = svd(W)
|
||||
B = U[:,:r] · √S[:r] # NO scaling pre-division
|
||||
A = √S[:r] · Vh[:r] # NO scaling pre-division
|
||||
W_res = W - scale · B·A # = W - scale · U·S·Vh (NOT
|
||||
exactly W - U·S·Vh)
|
||||
forward: W_res·x + scale · B·A·x # net: W·x at t=0 ✓
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ Bit-identity at t=0 holds for any α (algebra cancels symmetrically): both
|
||||
peft and ours satisfy W_res + scale·BA = W round-tripped.
|
||||
- ⚠ A/B starting magnitudes differ when α≠r. peft pre-divides S/scale so
|
||||
||B||,||A|| ∝ √(S/scale); ours has ||B||,||A|| ∝ √S. Update dynamics (α/r ·
|
||||
ΔB·A) therefore differ by a factor of √scale per matrix. For α=r (scale=1)
|
||||
they're identical — and the docstring already says "Pass alpha=r for
|
||||
paper-faithful scale=1".
|
||||
- ✓ Linear-only restriction is a deliberate scope cut.
|
||||
- ✓ cfg.dtype cast on A,B before storage matches peft's to(dtype) after save.
|
||||
- ✓ fp32 SVD then (W - scale·BA) in fp32 before bf16 cast — comment "fp32
|
||||
subtraction so W_res stays accurate" is honoured. peft does same via
|
||||
weight.to(torch.float32).
|
||||
|
||||
Verdict: correct for α = r. For α ≠ r, behaves differently from peft (still
|
||||
mathematically valid PiSSA, just different parameter scaling of A/B). Either
|
||||
fix (apply Sr /= scale like peft) or tighten the docstring to "α=r required"
|
||||
and assert it. Mild fix; not a correctness bug.
|
||||
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● 4. DeLoRA review
|
||||
|
||||
Paper (Bini et al. 2025, ICLR'25):
|
||||
W' = W + (λ/r) · B · Ξ · A, Ξ_ii = 1/(||b_i|| · ||a_i||)
|
||||
init: B=0 → ΔW=0
|
||||
default λ=15
|
||||
Crucial: w_norm is per-input-channel ||W||_2 (col-norm), captured ONCE at init
|
||||
(frozen reference) — not Frobenius scalar.
|
||||
|
||||
peft peft_delora_layer.py:226-264:
|
||||
forward(x):
|
||||
h = (x * w_norm) @ A.T # broadcast w_norm over
|
||||
input dim
|
||||
An = clamp(A.norm(dim=1), 1e-4)
|
||||
Bn = clamp(B.norm(dim=0), 1e-4)
|
||||
scale_r = (λ/r) / (An * Bn) # (r,)
|
||||
h = h * scale_r
|
||||
h = h @ B.T
|
||||
return base_out + h
|
||||
init: A ← kaiming(√5), B ← 0, λ ← 15.0, w_norm ← ||W||_2 col (frozen
|
||||
buffer)
|
||||
|
||||
ours src/lora_lite/variants/delora.py:55-81:
|
||||
forward(x, y):
|
||||
x_scaled = x * lora_wnorm
|
||||
h = einsum(x_scaled, A, "... i, r i -> ... r") # = (x * w_norm) @ A.T
|
||||
An = clamp(A.norm(dim=1), 1e-4)
|
||||
Bn = clamp(B.norm(dim=0), 1e-4)
|
||||
scale_r = (λ/r) / (An * Bn)
|
||||
h = h * scale_r
|
||||
delta = einsum(h, B, "... r, o r -> ... o") # = h @ B.T
|
||||
return y + delta
|
||||
init: A ← kaiming(√5), B ← 0, λ ← lambda0 (default 0.0, NOT 15.0),
|
||||
lora_wnorm ← ||W||_2 col (frozen buffer)
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ math line-for-line identical: same dim-0 col-norm, same per-rank A/B
|
||||
normalization, same gradient path (norms not detached, A,B receive gradient
|
||||
through them — matches peft exactly).
|
||||
- ✓ frozen lora_wnorm buffer (peft uses BufferDict, ours uses as_buffer=True).
|
||||
- ⚠ default lambda0=0.0 differs from peft's 15.0. Docstring is honest about
|
||||
this ("0.0 = bit-identity at t=0, but A,B get zero grad until lambda moves").
|
||||
At λ=0, gradients on A,B vanish since delta ∝ λ, so training stalls until λ
|
||||
moves off zero. This is documented but is a sharp foot-gun. peft's 15.0
|
||||
default avoids it.
|
||||
- ✓ even-r restriction not relevant here (HRA only).
|
||||
- ✓ nn.Linear-only restriction documented.
|
||||
|
||||
Verdict: correct. The lambda0 default is a deliberate "bit-identity but
|
||||
stalls" choice — keep, but consider noting in docstring that for actual
|
||||
training you'd want λ0≈1-15 (paper uses 15).
|
||||
|
||||
Searched for 1 pattern, read 1 file (ctrl+o to expand)
|
||||
|
||||
● 5. EVA review
|
||||
|
||||
Paper (Paischer et al. 2024):
|
||||
For each target layer l:
|
||||
Collect inputs x ∈ R^{N × d_in} on calibration batches
|
||||
Vh = right_singular_vectors(x) # top-r principal input dirs
|
||||
A_l ← Vh[:r] # data-driven init
|
||||
B_l ← 0
|
||||
forward = standard LoRA
|
||||
Optional: rank redistribution by explained variance (ρ-budget across layers)
|
||||
Optional: hash-dedup layers with identical inputs (Q/K/V share x)
|
||||
|
||||
peft peft_eva.py:
|
||||
For each calibration batch:
|
||||
IncrementalPCA.partial_fit(layer_input) # streaming,
|
||||
memory-efficient
|
||||
Hash inputs; dedup layers seeing same x.
|
||||
Track convergence via cossim(components_t, components_{t-1})
|
||||
After:
|
||||
Redistribute ranks by explained_variance_ratio (ρ scale)
|
||||
A_l ← svd.components_[:rank_l] # top-rank rows
|
||||
B_l ← 0
|
||||
forward = vanilla LoRA path (`scale·B@A·x`)
|
||||
|
||||
ours src/lora_lite/variants/eva.py:
|
||||
group_init(model, targets, calibration_data):
|
||||
hooks capture pre-forward inputs per layer
|
||||
for batch in calibration_data: model(batch)
|
||||
for each target:
|
||||
X = cat(captured) # (N, d_in), float32 on cpu
|
||||
_,_,Vh = torch.linalg.svd(X, full_matrices=False)
|
||||
A ← Vh[:r] # uniform r per layer
|
||||
forward(x, y):
|
||||
return y + (α/r) · (B @ A @ x) # B=0 at t=0 → identity
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ Top-r right singular vectors of pooled inputs → A. Math identical.
|
||||
- ✓ B=0 init → identity at t=0.
|
||||
- ✓ A,B both trainable post-init (peft same).
|
||||
- ✗ Full SVD on cpu-pooled tensor vs. peft's IncrementalPCA — limits us to
|
||||
small calibration sets (we cat all activations in RAM). The docstring
|
||||
acknowledges this.
|
||||
- ✗ No rank redistribution (rho param). All targets get same r. Acknowledged.
|
||||
- ✗ No QKV input dedup. Acknowledged. Cost: 3× SVD work for layers that share
|
||||
input. Correct, just slower.
|
||||
- ✗ No convergence check. Single pass over calibration_data. Acknowledged.
|
||||
- ⚠ Captured tensors moved to cpu+fp32 — fine, but a noisy 1k-batch on a 70B
|
||||
model = lots of RAM. Just a scaling note.
|
||||
|
||||
Verdict: correct minimal EVA, all simplifications stated up-front in
|
||||
docstring. No bugs. If you ever want to scale to large calibration corpora,
|
||||
swap the torch.cat → svd for IncrementalPCA.
|
||||
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● 6. HRA review
|
||||
|
||||
Paper (Yuan et al. 2024):
|
||||
W' = W·R, R = ∏_{i=0..r-1} H_i, H_i = I - 2·u_i·u_i^T / ||u_i||^2
|
||||
init (paper): independent random columns
|
||||
peft variant: paired columns (u_{2k}=u_{2k+1}) → R=I exactly at t=0, even r
|
||||
GS variant (apply_GS=True): explicit Gram-Schmidt for strict orthogonality
|
||||
Loss-side: orthogonality regularizer (Eq. 6); not part of layer.
|
||||
|
||||
peft peft_hra_layer.py:208-265:
|
||||
hra_u: (d_in, r) # COLUMNS are vectors
|
||||
get_delta_weight():
|
||||
opt_u = opt_u / opt_u.norm(dim=0) # pre-normalize columns
|
||||
R = I_{d_in}
|
||||
for i in 0..r-1: # forward order
|
||||
u = opt_u[:, i]
|
||||
R = R - 2·R·u·u^T # R ← R · H_i
|
||||
forward(x):
|
||||
R = product over active adapters
|
||||
return F.linear(x, W @ R, bias) # = x · R^T · W^T
|
||||
(effectively R applied to x)
|
||||
init: half_u = kaiming((d_in, r/2)); u = repeat_interleave(half_u, 2, dim=1)
|
||||
# cols 0=1, 2=3, ...
|
||||
|
||||
ours src/lora_lite/variants/hra.py:53-75:
|
||||
lora_U: (r, d_in) # ROWS are vectors
|
||||
(transpose convention)
|
||||
forward_input(x):
|
||||
Rx = x
|
||||
for i = r-1 downto 0: # reverse order, applied
|
||||
directly
|
||||
u = U[i]
|
||||
sq = (u·u).sum().clamp_min(1e-12)
|
||||
coeff = (Rx · u) · (2/sq)
|
||||
Rx = Rx - coeff · u # Rx ← Rx · H_i
|
||||
return Rx # then base layer applies W
|
||||
to this
|
||||
init: half = kaiming((r/2, d_in)); U = repeat_interleave(half, 2, dim=0) #
|
||||
rows 0=1, 2=3, ...
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ Math equivalent. peft computes R = H_0·H_1·...·H_{r-1} (col-vector
|
||||
convention), then applies via x·R^T = x·H_{r-1}·...·H_0 (since H is
|
||||
symmetric). Ours iterates i = r-1 → 0 doing Rx ← Rx·H_i, computing
|
||||
x·H_{r-1}·...·H_0. Same.
|
||||
- ✓ symmetric init: row pairs vs col pairs — consistent with our row-vector
|
||||
convention.
|
||||
- ✓ identity at t=0: paired u_a=u_b → H_a·H_a=I (involution). R=I.
|
||||
- ⚠ No apply_GS=True path. peft offers Gram-Schmidt orthogonalization.
|
||||
Docstring acknowledges via "Note: paper's orthogonality regularizer (Eq. 6) is
|
||||
loss-side". Ours uses pure Householder products (which already give exactly
|
||||
orthogonal R at any step), so GS is redundant unless you want unit-norm
|
||||
columns. Fine to omit.
|
||||
- ✓ More efficient than peft. Avoids materializing (d_in × d_in) R matrix;
|
||||
applies reflections directly to x in O(r·d_in) not O(r·d_in²).
|
||||
- ⚠ Slight gradient-flow difference. peft pre-normalizes opt_u /
|
||||
||opt_u||_col, so gradients flow through normalization. Ours keeps ||u||² in
|
||||
the denominator each step, gradients flow through that. Both define the same
|
||||
orthogonal H_i; gradients should give equivalent training to within numerical
|
||||
noise, but not bit-identical.
|
||||
- ✓ even-r requirement raised loudly (peft warns + falls back; we raise —
|
||||
preferable for fail-fast).
|
||||
|
||||
Verdict: correct + faster than peft. Ordering bug from 67a6daf is properly
|
||||
fixed (verified: i = r-1 → 0 iteration matches peft's R = H_0..H_{r-1} then
|
||||
x·R^T).
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● 7. IA3 review
|
||||
|
||||
Paper (Liu et al. 2022):
|
||||
y = ℓ ⊙ (W·x) for attention K, V (output-side gate)
|
||||
y = W_2·(ℓ_ff ⊙ γ(W_1·x)) for FF (mid-FFN gate, T5)
|
||||
init: ℓ = 1
|
||||
|
||||
peft peft_ia3_layer.py:161-188 (single class with is_feedforward flag):
|
||||
if is_feedforward:
|
||||
g shape (1, d_in)
|
||||
interm = x * g
|
||||
return base_layer(interm) # input-side
|
||||
else:
|
||||
g shape (d_out, 1)
|
||||
result = base_layer(x)
|
||||
return result * g # output-side
|
||||
init: g ← ones (when init_ia3_weights=True, default)
|
||||
default targets: ["k_proj", "v_proj", "down_proj"];
|
||||
feedforward_modules=["down_proj"]
|
||||
|
||||
ours src/lora_lite/variants/ia3.py:38-76 (two classes IA3, IA3FF):
|
||||
IA3 (output-side, e.g. k_proj/v_proj):
|
||||
g shape (d_out,); init ones
|
||||
forward(x, y) = y * g
|
||||
|
||||
IA3FF (input-side, e.g. down_proj/fc2):
|
||||
g shape (d_in,); init ones
|
||||
forward_input(x) = x * g # base_layer applied
|
||||
after by host
|
||||
|
||||
Diff vs peft:
|
||||
- ✓ math identical to peft's two branches.
|
||||
- ✓ init=ones matches.
|
||||
- ✓ Splitting into two classes (rather than is_feedforward flag) is cleaner —
|
||||
fail-fast: which class you register fixes the side, no boolean to misset.
|
||||
- ⚠ peft's output-side path also scales base_layer.bias during merge. We
|
||||
don't merge (forward-only adapter), so no issue at training time — but if you
|
||||
ever add merge, watch this.
|
||||
- ⚠ docstring example uses regex r"\\.k_proj$" which under YAML/Python would
|
||||
match a literal .k_proj string-end; check that target_names is treated as
|
||||
regex by your attach() (matches the behavior in variant.py).
|
||||
- ⚠ Position-on-Llama: peft applies "ff" gate at down_proj input, which is
|
||||
gate*up (post-SiLU). This deviates from paper's T5 position (post-activation
|
||||
of single W_1) but matches peft's default for Llama-style. Acknowledged via
|
||||
docstring example.
|
||||
|
||||
Verdict: correct, cleaner than peft. No bugs.
|
||||
|
||||
❯ ko! so what do we need to fix?
|
||||
|
||||
● Fixes needed (from reviews 1-7)
|
||||
|
||||
Real bug (1):
|
||||
- dora.py:62-63 — missing .detach() on the LoRA path inside the norm. Paper
|
||||
§4.3 explicitly says treat ||V+ΔV||_c as a constant; peft detaches twice.
|
||||
Without it, grads flow through the norm denominator → unstable training + ~2×
|
||||
memory.
|
||||
V = layer.weight + scale * BA.detach() # detach BA here
|
||||
v_norm = V.norm(dim=1).clamp_min(1e-12).detach() # belt & braces
|
||||
|
||||
Behavioral / docstring (2):
|
||||
- pissa.py — at α≠r, A,B starting magnitude differs from peft by √scale per
|
||||
matrix (peft pre-divides S/scale). Either match peft (Sr /= scale before
|
||||
computing √Sr factors) or assert cfg.alpha == cfg.r and tighten the docstring.
|
||||
- delora.py:36 — lambda0=0.0 default is bit-identity but a foot-gun: at λ=0
|
||||
the gradient on A,B is zero, so training is dead until λ moves. Either change
|
||||
default to peft's 15.0, or expand the existing comment to flag this for new
|
||||
users.
|
||||
|
||||
Clean (4): lora, eva, hra, ia3 — no fixes.
|
||||
|
||||
Pending: antipasto.py (review #8 not yet done; want me to continue?).
|
||||
|
||||
✻ Perused for 5m 30s
|
||||
|
||||
❯ sure do antipasto now...
|
||||
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● 8. AntiPaSTO review
|
||||
|
||||
Reference (antipasto3, Flax):
|
||||
kernel = U·diag(S)·Vh + W_res # kernel: (in, out)
|
||||
init U,S,Vh,W_res = svd(kernel) (frozen, fp32 / bf16 W_res)
|
||||
trainable:
|
||||
delta_s ∼ truncated_normal·4e-4 + 4e-4 # small POSITIVE bias
|
||||
(sym-break)
|
||||
rot_T ∼ truncated_normal·1e-4 # small random
|
||||
forward(x, alpha):
|
||||
A = scatter(rot_T into upper-triu of (n_blocks, bs, bs))
|
||||
A = 0.5·(A - A^T); A = a_lim·tanh(A/a_lim)
|
||||
X = α·A/2 # << α scales rotation
|
||||
R = (I-X)^-1 (I+X) # block Cayley
|
||||
if rotate_U: U_rot = U @ R # einsum, never builds full
|
||||
(r,r)
|
||||
if rotate_V: Vh_rot = R @ Vh
|
||||
S_scaled = S + α·delta_s
|
||||
return (x @ U_rot)·S_scaled @ Vh_rot + x @ W_res
|
||||
Identity property: bidirectional symmetry — R(α=+1) and R(α=-1) are exact
|
||||
inverses.
|
||||
At α=0: U,S,Vh untouched → kernel reproduced.
|
||||
|
||||
ours src/lora_lite/variants/antipasto.py:
|
||||
weight (pytorch): (d_out, d_in); SVD: U:(d_out,r), S:(r,), Vh:(r,d_in)
|
||||
init: U,S,Vh frozen buffers; W_res mutates layer.weight (PiSSA-style)
|
||||
trainable:
|
||||
delta_s = 0
|
||||
rot_T = 0 # << EXACT identity at t=0
|
||||
forward(x, y):
|
||||
A = scatter(rot_T) # (n_blocks, bs, bs)
|
||||
A = 0.5·(A - A^T); A = a_lim·tanh(A/a_lim)
|
||||
X = A/2 # << no α
|
||||
R_blocks = Cayley(X)
|
||||
R = block_diag(R_blocks) # << builds full (r,r) via
|
||||
for-loop
|
||||
if rotate_basis=="V": Vh_eff = R @ Vh
|
||||
elif rotate_basis=="U": U_eff = U @ R^T
|
||||
S_eff = S + delta_s
|
||||
h = (x @ Vh_eff^T) · S_eff
|
||||
delta = h @ U_eff^T
|
||||
return y + delta # y = x @ W_res^T already
|
||||
Identity at t=0: rot_T=0 → R=I, delta_s=0 → delta = x @ Vh^T @ diag(S) @ U^T =
|
||||
x @ W_svd^T → exact bit-identity.
|
||||
|
||||
Diff vs antipasto3:
|
||||
|
||||
- ⚠ No alpha steering coefficient. antipasto3's headline feature —
|
||||
bidirectional symmetry (R(+α)·R(-α)=I) — is dropped. If you intend antipasto
|
||||
as a fine-tuning-only adapter, this is a deliberate scope cut; if you ever
|
||||
want bidirectional steering at inference, you'd need to thread an α through
|
||||
the forward.
|
||||
- ⚠ Rotation chirality flipped. antipasto3 rotates the input basis by R (U @
|
||||
R); ours rotates by R^T (Vh_eff = R @ Vh ⇒ (x @ Vh^T) @ R^T). Same
|
||||
orthogonal-group element up to inversion, so gradient-equivalent under sign
|
||||
flip of rot_T. Not a correctness bug, but checkpoints aren't transferable
|
||||
between the two impls without a sign flip.
|
||||
- ✓ Stricter t=0 identity. Ours zero-inits both delta_s and rot_T → exact
|
||||
bit-identity. antipasto3 nudges delta_s ≈ +4e-4 and rot_T ∼ 1e-4 for symmetry
|
||||
breaking.
|
||||
Foot-gun: at exact zero rotation, ∂loss/∂rot_T may be sign-symmetric (rotating
|
||||
in either direction gives same loss to first order), so optimizer can stall
|
||||
on rot_T while delta_s carries the load. antipasto3's small bias avoids this.
|
||||
Worth a docstring note.
|
||||
- ✗ Perf: full (r,r) block-diagonal materialized via Python for-loop in
|
||||
_block_diag. antipasto3 uses einsum('dnb,nbc->dnc', U_reshaped, R_blocks) to
|
||||
apply each block directly without ever building the (r,r) matrix. For r=64,
|
||||
bs=4 this is 16 reflective blocks; the for-loop is O(r²) memory and Python
|
||||
overhead. Easy fix:
|
||||
# rotate V (input basis): replace block_diag + matmul with einsum
|
||||
Vh_blocks = Vh.reshape(n_blocks, bs, d_in)
|
||||
Vh_eff = einsum(R_blocks, Vh_blocks, "n a b, n b i -> n a i").reshape(r, d_in)
|
||||
- ⚠ rotate_basis enum ("V" XOR "U") vs. antipasto3's two booleans (can do
|
||||
both). Minor expressivity loss.
|
||||
- ✓ Cayley, skew construction, tanh angle clamping all match antipasto3
|
||||
exactly (incl. the 0.5 factor on (A - A^T)).
|
||||
- ✓ W_res via PiSSA-style in-place mutation of layer.weight is more
|
||||
memory-efficient than antipasto3's separate buffer; semantically identical.
|
||||
- ⚠ arxiv URL: https://arxiv.org/abs/2601.07473 — earlier I confirmed it
|
||||
returns HTTP 200, so the paper exists. Fine.
|
||||
|
||||
Verdict: correct as a fine-tuning adapter with stricter t=0 identity. Two
|
||||
divergences from antipasto3 worth deciding on:
|
||||
|
||||
1. Chirality: trivial sign-flip in skew if you want checkpoint compat with
|
||||
antipasto3 — otherwise leave.
|
||||
2. No α: if antipasto needs to support inference-time steering (the paper's
|
||||
selling point), thread an α scalar through the forward and S_eff = S +
|
||||
α·delta_s; X = α·A/2.
|
||||
|
||||
And one perf fix: kill _block_diag, use einsum.
|
||||
|
||||
❯ ok now make a task for each fix and fix pls
|
||||
|
||||
● Update(src/lora_lite/variants/dora.py)
|
||||
⎿ Added 4 lines, removed 3 lines
|
||||
57 ) -> Float[T, '*B o']:
|
||||
58 cfg = layer._lora_cfg
|
||||
59 scale = cfg.alpha / cfg.r
|
||||
60 - # V = W + scale * B @ A
|
||||
60 + # Paper §4.3: treat ||V+ΔV||_c as a constant (detach fr
|
||||
+om grad graph) for
|
||||
61 + # stability and ~2x lower memory. Match peft (lora_weig
|
||||
+ht.detach + weight_norm.detach).
|
||||
62 BA = einsum(layer.lora_B, layer.lora_A, "o r, r i -> o
|
||||
i")
|
||||
62 - V = layer.weight + scale * BA
|
||||
-# (d_out, d_in)
|
||||
63 - v_norm = V.norm(dim=1).clamp_min(1e-12)
|
||||
-# (d_out,)
|
||||
63 + V = layer.weight + scale * BA.detach()
|
||||
+# (d_out, d_in)
|
||||
64 + v_norm = V.norm(dim=1).clamp_min(1e-12).detach()
|
||||
+# (d_out,)
|
||||
65 # Bias passes through unscaled (matches peft).
|
||||
66 bias = getattr(layer, "bias", None)
|
||||
67 wx = y if bias is None else (y - bias)
|
||||
|
||||
● Update(src/lora_lite/variants/pissa.py)
|
||||
⎿ Added 4 lines, removed 3 lines
|
||||
3 Meng et al. 2024 https://arxiv.org/abs/2404.02948
|
||||
4
|
||||
5 W = U S Vh (truncated to top-r)
|
||||
6 - B = U sqrt(S), A = sqrt(S) Vh, W_res = W - B A
|
||||
6 + Sr_eff = Sr / (alpha/r) # peft-sty
|
||||
+le: pre-divide so A/B
|
||||
7 + B = U sqrt(Sr_eff), A = sqrt(Sr_eff) Vh # update d
|
||||
+ynamics match for any alpha
|
||||
8 + W_res = W - (alpha/r) B A = W - U Sr Vh # scaling
|
||||
+cancels symmetrically
|
||||
9
|
||||
8 -Identity at t=0: W_res + B@A == W (bf16 round-trip, not bit-exa
|
||||
-ct).
|
||||
9 -Pass alpha=r for paper-faithful scale=1.
|
||||
10 +Identity at t=0: W_res + (alpha/r) B@A == W (fp32 round-trip, b
|
||||
+f16 cast can drift).
|
||||
11
|
||||
12 Refs:
|
||||
13 - paper: https://github.com/MuLabPKU/PiSSA/blob/main/utils/in
|
||||
it_pissa.py
|
||||
|
||||
● Update(src/lora_lite/variants/pissa.py)
|
||||
⎿ Added 7 lines, removed 3 lines
|
||||
52 W = layer.weight.data.float() # (
|
||||
d_out, d_in)
|
||||
53 U, S, Vh = torch.linalg.svd(W, full_matrices=False)
|
||||
54 r = cfg.r
|
||||
55 + scale = cfg.alpha / cfg.r
|
||||
56 Ur, Sr, Vhr = U[:, :r], S[:r], Vh[:r, :]
|
||||
56 - sqrtS = Sr.sqrt()
|
||||
57 - # B @ A = Ur diag(Sr) Vhr; pick B = Ur sqrt(Sr), A =
|
||||
-sqrt(Sr) * Vhr
|
||||
57 + # Pre-divide Sr by scaling so A/B carry "natural" magni
|
||||
+tudes for any alpha
|
||||
58 + # (peft pissa_init does this; needed so the alpha/r sca
|
||||
+ling on the forward
|
||||
59 + # gives matched update dynamics rather than re-scaling
|
||||
+A,B by sqrt(scale)).
|
||||
60 + Sr_eff = Sr / scale
|
||||
61 + sqrtS = Sr_eff.sqrt()
|
||||
62 + # B @ A = Ur diag(Sr/scale) Vhr; W_res = W - scale * B
|
||||
+@A = W - Ur diag(Sr) Vhr.
|
||||
63 B = (Ur * sqrtS).to(cfg.dtype)
|
||||
64 A = (sqrtS[:, None] * Vhr).to(cfg.dtype)
|
||||
65 layer.lora_B.data.copy_(B)
|
||||
66 layer.lora_A.data.copy_(A)
|
||||
67 # fp32 subtraction so W_res stays accurate.
|
||||
68 BA = (B.float() @ A.float())
|
||||
64 - scale = cfg.alpha / cfg.r
|
||||
69 layer.weight.data.copy_((W - scale * BA).to(layer.weigh
|
||||
t.dtype))
|
||||
70
|
||||
71 @staticmethod
|
||||
● Update(src/lora_lite/variants/antipasto.py)
|
||||
⎿ Removed 9 lines
|
||||
59 return _cayley(A)
|
||||
60
|
||||
61
|
||||
62 -def _block_diag(blocks: torch.Tensor) -> torch.Tensor:
|
||||
63 - """(n_blocks, bs, bs) -> (n_blocks*bs, n_blocks*bs) block-d
|
||||
-iagonal."""
|
||||
64 - n, bs, _ = blocks.shape
|
||||
65 - out = blocks.new_zeros(n * bs, n * bs)
|
||||
66 - for i in range(n):
|
||||
67 - out[i * bs : (i + 1) * bs, i * bs : (i + 1) * bs] = blo
|
||||
-cks[i]
|
||||
68 - return out
|
||||
69 -
|
||||
70 -
|
||||
62 @register
|
||||
63 class AntiPaSTO:
|
||||
64 name = "antipasto"
|
||||
|
||||
● Update(src/lora_lite/variants/antipasto.py)
|
||||
⎿ Added 11 lines, removed 4 lines
|
||||
114 S = layer.lora_S.to(x.dtype)
|
||||
# (r,)
|
||||
115 Vh = layer.lora_Vh.to(x.dtype)
|
||||
# (r, d_in)
|
||||
116
|
||||
117 - R_blocks = _build_rotation(layer.lora_rot_T.float(), b
|
||||
-s, max_angle)
|
||||
118 - R = _block_diag(R_blocks).to(x.dtype)
|
||||
-# (r, r)
|
||||
117 + R_blocks = _build_rotation(layer.lora_rot_T.float(), b
|
||||
+s, max_angle).to(x.dtype)
|
||||
118 + n_blocks, _, _ = R_blocks.shape
|
||||
+# (n, bs, bs)
|
||||
119 + d_in, d_out = Vh.shape[1], U.shape[0]
|
||||
120 + r = n_blocks * bs
|
||||
121
|
||||
122 + # Apply block-diagonal R per-block via einsum, never m
|
||||
+aterializing (r,r).
|
||||
123 if rotate_basis == "V":
|
||||
121 - Vh_eff = R @ Vh
|
||||
-# rotate INPUT basis
|
||||
124 + # Vh_eff = R @ Vh, viewed block-wise on the r-axis
|
||||
+.
|
||||
125 + Vh_blocks = Vh.reshape(n_blocks, bs, d_in)
|
||||
126 + Vh_eff = einsum(R_blocks, Vh_blocks, "n a b, n b i
|
||||
+ -> n a i").reshape(r, d_in)
|
||||
127 U_eff = U
|
||||
128 elif rotate_basis == "U":
|
||||
129 + # U_eff = U @ R.T, viewed block-wise on the r-axis
|
||||
+.
|
||||
130 + U_blocks = U.reshape(d_out, n_blocks, bs)
|
||||
131 + U_eff = einsum(U_blocks, R_blocks, "d n b, n c b -
|
||||
+> d n c").reshape(d_out, r)
|
||||
132 Vh_eff = Vh
|
||||
125 - U_eff = U @ R.T
|
||||
-# rotate OUTPUT basis
|
||||
133 else:
|
||||
134 raise ValueError(f"rotate_basis must be 'U' or 'V'
|
||||
, got {rotate_basis!r}")
|
||||
135
|
||||
|
||||
● Verify the antipasto change with a quick import + a smoke equivalence check.
|
||||
|
||||
● Bash(python -c "
|
||||
import torch…)
|
||||
⎿ Error: Exit code 127
|
||||
(eval):1: command not found: python
|
||||
|
||||
❯ > einsum(R_blocks, Vh_blocks, "n a b, n b i -> n a i").reshape(r, d_in)
|
||||
|
||||
should not the reshape use einops reshape? there is also reduce
|
||||
|
||||
● You're right — use einops rearrange for the shape ops to keep them
|
||||
self-documenting. Updating.
|
||||
|
||||
Read 1 file (ctrl+o to expand)
|
||||
|
||||
● Update(src/lora_lite/variants/antipasto.py)
|
||||
⎿ Added 1 line, removed 1 line
|
||||
19 from typing import Literal
|
||||
20
|
||||
21 import torch
|
||||
22 -from einops import einsum
|
||||
22 +from einops import einsum, rearrange
|
||||
23 from jaxtyping import Float
|
||||
24 from torch import nn, Tensor as T
|
||||
25
|
||||
|
||||
● Update(src/lora_lite/variants/antipasto.py)
|
||||
⎿ Added 7 lines, removed 7 lines
|
||||
115 Vh = layer.lora_Vh.to(x.dtype)
|
||||
# (r, d_in)
|
||||
116
|
||||
117 R_blocks = _build_rotation(layer.lora_rot_T.float(), b
|
||||
s, max_angle).to(x.dtype)
|
||||
118 - n_blocks, _, _ = R_blocks.shape
|
||||
-# (n, bs, bs)
|
||||
119 - d_in, d_out = Vh.shape[1], U.shape[0]
|
||||
120 - r = n_blocks * bs
|
||||
118 + n_blocks = R_blocks.shape[0]
|
||||
+# R_blocks: (n, bs, bs)
|
||||
119
|
||||
120 # Apply block-diagonal R per-block via einsum, never m
|
||||
aterializing (r,r).
|
||||
121 if rotate_basis == "V":
|
||||
122 # Vh_eff = R @ Vh, viewed block-wise on the r-axis
|
||||
.
|
||||
125 - Vh_blocks = Vh.reshape(n_blocks, bs, d_in)
|
||||
126 - Vh_eff = einsum(R_blocks, Vh_blocks, "n a b, n b i
|
||||
- -> n a i").reshape(r, d_in)
|
||||
123 + Vh_blocks = rearrange(Vh, "(n a) i -> n a i", n=n_
|
||||
+blocks)
|
||||
124 + Vh_rot = einsum(R_blocks, Vh_blocks, "n a b, n b i
|
||||
+ -> n a i")
|
||||
125 + Vh_eff = rearrange(Vh_rot, "n a i -> (n a) i")
|
||||
126 U_eff = U
|
||||
127 elif rotate_basis == "U":
|
||||
128 # U_eff = U @ R.T, viewed block-wise on the r-axis
|
||||
.
|
||||
130 - U_blocks = U.reshape(d_out, n_blocks, bs)
|
||||
131 - U_eff = einsum(U_blocks, R_blocks, "d n b, n c b -
|
||||
-> d n c").reshape(d_out, r)
|
||||
129 + U_blocks = rearrange(U, "d (n b) -> d n b", n=n_bl
|
||||
+ocks)
|
||||
130 + U_rot = einsum(U_blocks, R_blocks, "d n b, n c b -
|
||||
+> d n c")
|
||||
131 + U_eff = rearrange(U_rot, "d n c -> d (n c)")
|
||||
132 Vh_eff = Vh
|
||||
133 else:
|
||||
134 raise ValueError(f"rotate_basis must be 'U' or 'V'
|
||||
, got {rotate_basis!r}")
|
||||
|
||||
● Now run the equivalence smoke check.
|
||||
|
||||
Listed 1 directory (ctrl+o to expand)
|
||||
|
||||
● Bash(.venv/bin/python -c "
|
||||
import torch…)
|
||||
⎿ t=0 identity err (V): 1.91e-06
|
||||
einsum vs blockdiag (V): 1.31e-06
|
||||
einsum vs blockdiag (U): 1.07e-06
|
||||
|
||||
● All three checks pass at fp32 round-trip noise (~1e-6).
|
||||
|
||||
● Quick sanity check on dora and pissa too.
|
||||
|
||||
● Bash(.venv/bin/python -c "
|
||||
import torch…)
|
||||
⎿ DoRA t=0 identity err: 0.00e+00
|
||||
DoRA grad on A: 0.3232 (should be > 0)
|
||||
DoRA grad on m: 3.2282 (should be > 0)
|
||||
… +4 lines (ctrl+o to expand)
|
||||
|
||||
● All three fixes verified:
|
||||
|
||||
- DoRA: t=0 bit-identity (err=0), grads still flow on A and m, but no longer
|
||||
through v_norm (paper §4.3 / peft).
|
||||
- PiSSA: bit-identity now holds for α ∈ {r/2, r, 2r}, not just α=r. Old impl
|
||||
was only correct at α=r.
|
||||
- AntiPaSTO: einsum path matches block-diag matmul to fp32 round-trip (~1e-6);
|
||||
t=0 bit-identity preserved.
|
||||
|
||||
Summary
|
||||
|
||||
┌───────────────────────────────────────────┬─────────────────────┬────────┐
|
||||
│ Fix │ File │ Status │
|
||||
├───────────────────────────────────────────┼─────────────────────┼────────┤
|
||||
│ 1. detach BA + v_norm (paper §4.3) │ dora.py:62-64 │ ✓ │
|
||||
├───────────────────────────────────────────┼─────────────────────┼────────┤
|
||||
│ 2. pre-divide Sr by scale to match peft │ pissa.py:51-65 │ ✓ │
|
||||
│ for any α │ │ │
|
||||
├───────────────────────────────────────────┼─────────────────────┼────────┤
|
||||
│ 3. einops rearrange + per-block einsum, │ antipasto.py:41-130 │ ✓ │
|
||||
│ dropped _block_diag │ │ │
|
||||
└───────────────────────────────────────────┴─────────────────────┴────────┘
|
||||
|
||||
Skipped (judgment calls — flag if you want them):
|
||||
- delora lambda0=0.0 default → already documented in docstring
|
||||
- antipasto α steering coefficient → design decision, ask before adding
|
||||
- antipasto rotation chirality vs antipasto3 → gradient-equivalent under sign
|
||||
flip; no checkpoint compat path anyway (jax↔torch)
|
||||
|
||||
✻ Tinkered for 5m 55s
|
||||
Reference in New Issue
Block a user