even more stuff

This commit is contained in:
2026-05-21 05:55:22 +00:00
parent 3fdb9f008b
commit 261fe5e43e
2 changed files with 920 additions and 0 deletions

View File

@@ -0,0 +1,419 @@
"""
DeepSeek-V4 CSA/HCA compressor kernels for CUTLASS CuTe DSL / Blackwell.
This is a production-oriented fusion boundary for the PyTorch reference in
`Pasted markdown.md`:
1. Run the projection stage as one or two packed Blackwell GEMMs:
CSA main: H @ [W_a_KV | W_a_Z | W_b_KV | W_b_Z] -> (N, 4*C)
CSA indexer: H @ [W_I_a_KV | W_I_a_Z | W_I_b_KV | W_I_b_Z] -> (N, 4*C_I)
HCA: H @ [W_KV | W_Z] -> (N, 2*C)
Use your tcgen05 / NVFP4 blockscaled GEMM here. Keep the packed projection
outputs in BF16/FP32 as you prefer; the compressor reads them as tensor
elements and accumulates the softmax reduction in FP32.
2. These native CuTe DSL kernels fuse:
bias add + column-wise softmax over token positions + weighted C sum
+ partial RoPE for KV output.
The full projection+compression single-kernel tcgen05 variant is possible, but it
needs your exact NVFP4 weight/scale layout and preferred tile shape. This file is
therefore the safe fusion seam: the expensive D x C math stays in your Blackwell
GEMM, while the small-position softmax/reduction/rope path avoids PyTorch ops and
extra materialization after projection.
Target dimensions used by DeepSeek-V4-Pro reference:
D=7168, C=512, C_I=128, CSA_M=4, HCA_M=128, NOPE=448, ROPE=64.
Assumptions:
* Tensors are contiguous row-major from PyTorch/DLPack.
* Projection buffers are laid out as described above.
* One sequence at a time. State/tail management remains on the caller side.
* For CSA continuation across calls, provide external previous-block B-side
projections for the first committed block. For fresh prefill, set
has_external_prev=False.
NOTE: I cannot compile this in this sandbox because CUTLASS CuTe DSL is not
installed here. It follows the CuTe DSL @jit/@kernel launch style, but you may
need tiny API-name edits if you are pinned to a specific CUTLASS 4.x commit.
"""
from __future__ import annotations
import torch
import cutlass
import cutlass.cute as cute
import cuda.bindings.driver as cuda
# -----------------------------------------------------------------------------
# Small helpers
# -----------------------------------------------------------------------------
LOG2_E = 1.44269504088896340736
def _ceil_div(a: int, b: int) -> int:
return (a + b - 1) // b
@cute.jit
def _expf(x: cutlass.Float32) -> cutlass.Float32:
# CuTe DSL exposes exp2 in cute.math; exp(x) = exp2(x * log2(e)).
return cute.math.exp2(x * cutlass.Float32(LOG2_E))
@cute.jit
def _read_csa_current_packed(
proj: cute.Tensor,
token_idx: cutlass.Int64,
col: cutlass.Int32,
kind: cutlass.Constexpr, # 0 Ca, 1 Za, 2 Cb, 3 Zb
OUT: cutlass.Constexpr,
) -> cutlass.Float32:
return proj[token_idx, kind * OUT + col].to(cutlass.Float32)
@cute.jit
def _read_hca_packed(
proj: cute.Tensor,
token_idx: cutlass.Int64,
col: cutlass.Int32,
kind: cutlass.Constexpr, # 0 C, 1 Z
C: cutlass.Constexpr,
) -> cutlass.Float32:
return proj[token_idx, kind * C + col].to(cutlass.Float32)
# -----------------------------------------------------------------------------
# CSA fused compressor from packed projections
# -----------------------------------------------------------------------------
@cute.jit
def _csa_raw_reduce_one_col(
proj: cute.Tensor, # (N_tokens, 4*OUT): Ca, Za, Cb, Zb
prev_b_proj: cute.Tensor, # (M, 2*OUT): Cb_prev, Zb_prev for first block, may be dummy
B_a: cute.Tensor, # (M, OUT)
B_b: cute.Tensor, # (M, OUT)
block_i: cutlass.Int64,
col: cutlass.Int32,
start_token_in_proj: cutlass.Int64,
has_external_prev: cutlass.Constexpr,
M: cutlass.Constexpr,
OUT: cutlass.Constexpr,
) -> cutlass.Float32:
"""Column-wise CSA softmax+weighted-sum for either main KV or indexer.
This implements exactly:
softmax([Z_a_cur + B_a ; Z_b_prev + B_b], dim=position)
sum S_a*C_a + sum S_b*C_b
with the block-0 no-prev case reducing over M current positions only.
"""
# First pass: max logit.
max_logit = cutlass.Float32(-3.4028234663852886e38)
# Current/current-a side always exists.
for p in cutlass.range_constexpr(M):
tok = start_token_in_proj + block_i * M + p
za = _read_csa_current_packed(proj, tok, col, 1, OUT) + B_a[p, col].to(cutlass.Float32)
max_logit = cute.math.fmax(max_logit, za)
# Previous/b side exists if this is not the first fresh block.
use_prev = (block_i > 0) or has_external_prev
if use_prev:
for p in cutlass.range_constexpr(M):
if block_i > 0:
tok_prev = start_token_in_proj + (block_i - 1) * M + p
zb = _read_csa_current_packed(proj, tok_prev, col, 3, OUT)
else:
# External previous block is packed as [Cb_prev | Zb_prev]
zb = prev_b_proj[p, OUT + col].to(cutlass.Float32)
zb = zb + B_b[p, col].to(cutlass.Float32)
max_logit = cute.math.fmax(max_logit, zb)
# Second pass: exp denominator and weighted value.
denom = cutlass.Float32(0.0)
acc = cutlass.Float32(0.0)
for p in cutlass.range_constexpr(M):
tok = start_token_in_proj + block_i * M + p
za = _read_csa_current_packed(proj, tok, col, 1, OUT) + B_a[p, col].to(cutlass.Float32)
ca = _read_csa_current_packed(proj, tok, col, 0, OUT)
e = _expf(za - max_logit)
denom += e
acc += e * ca
if use_prev:
for p in cutlass.range_constexpr(M):
if block_i > 0:
tok_prev = start_token_in_proj + (block_i - 1) * M + p
cb = _read_csa_current_packed(proj, tok_prev, col, 2, OUT)
zb = _read_csa_current_packed(proj, tok_prev, col, 3, OUT)
else:
cb = prev_b_proj[p, col].to(cutlass.Float32)
zb = prev_b_proj[p, OUT + col].to(cutlass.Float32)
zb = zb + B_b[p, col].to(cutlass.Float32)
e = _expf(zb - max_logit)
denom += e
acc += e * cb
return acc / denom
@cute.kernel
def csa_compress_projected_kernel(
proj_main: cute.Tensor, # (N_tokens, 4*C)
proj_indexer: cute.Tensor, # (N_tokens, 4*C_I)
prev_main_b: cute.Tensor, # (M, 2*C), Cb_prev|Zb_prev; dummy if no ext prev
prev_indexer_b: cute.Tensor, # (M, 2*C_I), Cb_prev|Zb_prev; dummy if no ext prev
B_a: cute.Tensor, # (M, C)
B_b: cute.Tensor, # (M, C)
B_I_a: cute.Tensor, # (M, C_I)
B_I_b: cute.Tensor, # (M, C_I)
cos_sin_cache: cute.Tensor, # (max_pos, ROPE), cos first half, sin second half
kv_out: cute.Tensor, # (n_blocks, C)
indexer_out: cute.Tensor, # (n_blocks, C_I)
n_blocks: cutlass.Int64,
start_token_in_proj: cutlass.Int64,
start_abs_pos: cutlass.Int64,
has_external_prev: cutlass.Constexpr,
M: cutlass.Constexpr,
C: cutlass.Constexpr,
C_I: cutlass.Constexpr,
NOPE: cutlass.Constexpr,
ROPE: cutlass.Constexpr,
COLS_PER_CTA: cutlass.Constexpr,
):
tx, _, _ = cute.arch.thread_idx()
bx, by, bz = cute.arch.block_idx()
block_i = bx.to(cutlass.Int64)
base_col = by * COLS_PER_CTA
col = base_col + tx
# bz == 0: main KV output with RoPE. bz == 1: indexer output, no RoPE.
if bz == 0:
if block_i < n_blocks and col < C:
# RoPE dims need even/odd pair. Let even lane compute/store both.
if col < NOPE:
val = _csa_raw_reduce_one_col(
proj_main, prev_main_b, B_a, B_b,
block_i, col, start_token_in_proj, has_external_prev, M, C,
)
kv_out[block_i, col] = val.to(kv_out.element_type)
else:
rope_col = col - NOPE
if (rope_col % 2) == 0:
# Compute pair and rotate by block end position.
x0 = _csa_raw_reduce_one_col(
proj_main, prev_main_b, B_a, B_b,
block_i, col, start_token_in_proj, has_external_prev, M, C,
)
x1 = _csa_raw_reduce_one_col(
proj_main, prev_main_b, B_a, B_b,
block_i, col + 1, start_token_in_proj, has_external_prev, M, C,
)
block_end_pos = start_abs_pos + block_i * M + (M - 1)
half_idx = rope_col // 2
cosv = cos_sin_cache[block_end_pos, half_idx].to(cutlass.Float32)
sinv = cos_sin_cache[block_end_pos, half_idx + ROPE // 2].to(cutlass.Float32)
kv_out[block_i, col] = (x0 * cosv - x1 * sinv).to(kv_out.element_type)
kv_out[block_i, col + 1] = (x0 * sinv + x1 * cosv).to(kv_out.element_type)
else:
if block_i < n_blocks and col < C_I:
val_i = _csa_raw_reduce_one_col(
proj_indexer, prev_indexer_b, B_I_a, B_I_b,
block_i, col, start_token_in_proj, has_external_prev, M, C_I,
)
indexer_out[block_i, col] = val_i.to(indexer_out.element_type)
@cute.jit
def launch_csa_compress_projected(
proj_main: cute.Tensor,
proj_indexer: cute.Tensor,
prev_main_b: cute.Tensor,
prev_indexer_b: cute.Tensor,
B_a: cute.Tensor,
B_b: cute.Tensor,
B_I_a: cute.Tensor,
B_I_b: cute.Tensor,
cos_sin_cache: cute.Tensor,
kv_out: cute.Tensor,
indexer_out: cute.Tensor,
n_blocks: int,
start_token_in_proj: int,
start_abs_pos: int,
has_external_prev: cutlass.Constexpr,
stream: cuda.CUstream,
M: cutlass.Constexpr = 4,
C: cutlass.Constexpr = 512,
C_I: cutlass.Constexpr = 128,
NOPE: cutlass.Constexpr = 448,
ROPE: cutlass.Constexpr = 64,
COLS_PER_CTA: cutlass.Constexpr = 128,
):
grid_y = _ceil_div(C, COLS_PER_CTA) # enough for main; indexer just masks col < C_I
csa_compress_projected_kernel(
proj_main,
proj_indexer,
prev_main_b,
prev_indexer_b,
B_a,
B_b,
B_I_a,
B_I_b,
cos_sin_cache,
kv_out,
indexer_out,
n_blocks,
start_token_in_proj,
start_abs_pos,
has_external_prev,
M,
C,
C_I,
NOPE,
ROPE,
COLS_PER_CTA,
).launch(
grid=[n_blocks, grid_y, 2],
block=[COLS_PER_CTA, 1, 1],
stream=stream,
)
# -----------------------------------------------------------------------------
# HCA fused compressor from packed projections
# -----------------------------------------------------------------------------
@cute.jit
def _hca_raw_reduce_one_col(
proj: cute.Tensor, # (N_tokens, 2*C): C, Z
B: cute.Tensor, # (M, C)
block_i: cutlass.Int64,
col: cutlass.Int32,
start_token_in_proj: cutlass.Int64,
M: cutlass.Constexpr,
C: cutlass.Constexpr,
) -> cutlass.Float32:
max_logit = cutlass.Float32(-3.4028234663852886e38)
for p in cutlass.range_constexpr(M):
tok = start_token_in_proj + block_i * M + p
z = _read_hca_packed(proj, tok, col, 1, C) + B[p, col].to(cutlass.Float32)
max_logit = cute.math.fmax(max_logit, z)
denom = cutlass.Float32(0.0)
acc = cutlass.Float32(0.0)
for p in cutlass.range_constexpr(M):
tok = start_token_in_proj + block_i * M + p
z = _read_hca_packed(proj, tok, col, 1, C) + B[p, col].to(cutlass.Float32)
c = _read_hca_packed(proj, tok, col, 0, C)
e = _expf(z - max_logit)
denom += e
acc += e * c
return acc / denom
@cute.kernel
def hca_compress_projected_kernel(
proj: cute.Tensor, # (N_tokens, 2*C)
B: cute.Tensor, # (M, C)
cos_sin_cache: cute.Tensor, # (max_pos, ROPE)
kv_out: cute.Tensor, # (n_blocks, C)
n_blocks: cutlass.Int64,
start_token_in_proj: cutlass.Int64,
start_abs_pos: cutlass.Int64,
M: cutlass.Constexpr,
C: cutlass.Constexpr,
NOPE: cutlass.Constexpr,
ROPE: cutlass.Constexpr,
COLS_PER_CTA: cutlass.Constexpr,
):
tx, _, _ = cute.arch.thread_idx()
bx, by, _ = cute.arch.block_idx()
block_i = bx.to(cutlass.Int64)
col = by * COLS_PER_CTA + tx
if block_i < n_blocks and col < C:
if col < NOPE:
val = _hca_raw_reduce_one_col(proj, B, block_i, col, start_token_in_proj, M, C)
kv_out[block_i, col] = val.to(kv_out.element_type)
else:
rope_col = col - NOPE
if (rope_col % 2) == 0:
x0 = _hca_raw_reduce_one_col(proj, B, block_i, col, start_token_in_proj, M, C)
x1 = _hca_raw_reduce_one_col(proj, B, block_i, col + 1, start_token_in_proj, M, C)
block_end_pos = start_abs_pos + block_i * M + (M - 1)
half_idx = rope_col // 2
cosv = cos_sin_cache[block_end_pos, half_idx].to(cutlass.Float32)
sinv = cos_sin_cache[block_end_pos, half_idx + ROPE // 2].to(cutlass.Float32)
kv_out[block_i, col] = (x0 * cosv - x1 * sinv).to(kv_out.element_type)
kv_out[block_i, col + 1] = (x0 * sinv + x1 * cosv).to(kv_out.element_type)
@cute.jit
def launch_hca_compress_projected(
proj: cute.Tensor,
B: cute.Tensor,
cos_sin_cache: cute.Tensor,
kv_out: cute.Tensor,
n_blocks: int,
start_token_in_proj: int,
start_abs_pos: int,
stream: cuda.CUstream,
M: cutlass.Constexpr = 128,
C: cutlass.Constexpr = 512,
NOPE: cutlass.Constexpr = 448,
ROPE: cutlass.Constexpr = 64,
COLS_PER_CTA: cutlass.Constexpr = 128,
):
grid_y = _ceil_div(C, COLS_PER_CTA)
hca_compress_projected_kernel(
proj,
B,
cos_sin_cache,
kv_out,
n_blocks,
start_token_in_proj,
start_abs_pos,
M,
C,
NOPE,
ROPE,
COLS_PER_CTA,
).launch(
grid=[n_blocks, grid_y, 1],
block=[COLS_PER_CTA, 1, 1],
stream=stream,
)
# -----------------------------------------------------------------------------
# PyTorch-side packing helpers
# -----------------------------------------------------------------------------
def pack_csa_main_weights(W_a_KV, W_a_Z, W_b_KV, W_b_Z):
"""Return W packed as [W_a_KV | W_a_Z | W_b_KV | W_b_Z]."""
return torch.cat([W_a_KV, W_a_Z, W_b_KV, W_b_Z], dim=1).contiguous()
def pack_csa_indexer_weights(W_I_a_KV, W_I_a_Z, W_I_b_KV, W_I_b_Z):
"""Return W_I packed as [W_I_a_KV | W_I_a_Z | W_I_b_KV | W_I_b_Z]."""
return torch.cat([W_I_a_KV, W_I_a_Z, W_I_b_KV, W_I_b_Z], dim=1).contiguous()
def pack_hca_weights(W_KV, W_Z):
"""Return W packed as [W_KV | W_Z]."""
return torch.cat([W_KV, W_Z], dim=1).contiguous()
def make_dummy_prev_b(M: int, OUT: int, *, device, dtype):
"""Dummy external previous-block projection for fresh prefill."""
return torch.empty((M, 2 * OUT), device=device, dtype=dtype)

View File

@@ -0,0 +1,501 @@
"""
mHC (Manifold-Constrained Hyper-Connections) — Inference Layer.
Implements Section 2.2 of the DeepSeek-V4 paper for the forward pass only.
At inference the Sinkhorn-Knopp constraint has already been enforced during
training, but B_l is still *dynamically generated* per-token from the input
residual state. So we still need to:
1. Project the flattened residual → raw A/B/C parameter values.
2. Apply sigmoid (A, C) and Sinkhorn-Knopp 20 iters (B).
3. Mix residual streams.
The only thing that changes vs training is that we skip the loss and gradient
through the Sinkhorn projection — the forward arithmetic is identical.
---------------------------------------------------------------------
V4-Pro reference dimensions (Section 4.2.1)
---------------------------------------------------------------------
d = 7168 hidden dim
n_hc = 4 hyper-connection expansion factor
N_proj = 24 fused output of W_pre(4) + W_res(16) + W_post(4)
K_proj = 4*7168 = 28672 = n_hc * d (flattened residual)
t_max = 20 Sinkhorn iterations
---------------------------------------------------------------------
Kernel dependency
---------------------------------------------------------------------
tf32_hc_prenorm_gemm (DeepGEMM, SM90/SM100)
a: (T, K) BF16 — flattened residual X_flat
b: (N, K) FP32 — stacked weight [W_pre; W_res; W_post]
d: (S, T, N) or (T, N) FP32 — raw projection outputs (pre-normalised)
sqr_sum: (S, T) or (T,) FP32 — Σ a² per token (for RMSNorm denominator)
num_splits = S (16 recommended for K=28672)
After the call:
d = d.sum(0) → (T, N)
sqr_sum = sqr_sum.sum(0) → (T,)
rms_scale = sqrt(K / (sqr_sum + eps))
d_norm = d * rms_scale[:,None] — equivalent to RMSNorm(X_flat) @ W_stacked
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Optional, Tuple
import torch
import torch.nn.functional as F
# ---------------------------------------------------------------------------
# Try importing DeepGEMM; fall back to plain BF16 matmul if unavailable.
# ---------------------------------------------------------------------------
try:
import deep_gemm
_HAS_DEEP_GEMM = True
except ImportError:
_HAS_DEEP_GEMM = False
NUM_SPLITS = 16 # K-split count for tf32_hc_prenorm_gemm numerical stability
EPS_RMSN = 1e-6
# ---------------------------------------------------------------------------
# Sinkhorn-Knopp projection (T batched 4×4 matrices, 20 iters)
# ---------------------------------------------------------------------------
def sinkhorn_knopp(
M: torch.Tensor, # (T, n, n) positive (after exp)
t_max: int = 20,
) -> torch.Tensor:
"""
Project each (n×n) positive matrix onto the Birkhoff polytope
(doubly stochastic matrices) via alternating row/col normalisation.
Paper eq. (8): M^(t) = T_r( T_c( M^(t-1) ) )
where T_r = row-normalise, T_c = col-normalise.
For n=4 and t_max=20 this is ~160 tiny operations — no kernel needed.
All ops stay on GPU via standard PyTorch.
"""
for _ in range(t_max):
M = M / (M.sum(dim=-1, keepdim=True) + EPS_RMSN) # T_r (row)
M = M / (M.sum(dim=-2, keepdim=True) + EPS_RMSN) # T_c (col)
return M
# ---------------------------------------------------------------------------
# Context carried between pre_block and post_block
# ---------------------------------------------------------------------------
@dataclass
class mHCContext:
"""Holds the per-token mixing matrices computed in pre_block."""
B_l: torch.Tensor # (T, n_hc, n_hc) doubly stochastic residual transform
C_l: torch.Tensor # (T, n_hc) output mapping (before unsqueeze)
# ---------------------------------------------------------------------------
# mHC layer
# ---------------------------------------------------------------------------
class mHCLayer:
"""
Wraps one transformer sub-layer (attention *or* MoE) with the mHC
residual update.
Typical call pattern per layer:
x_in, ctx = mhc.pre_block(X_l)
F_out = transformer_sublayer(x_in) # (T, d)
X_next = mhc.post_block(X_l, F_out, ctx)
where X_l has shape (T, n_hc, d) — the expanded residual state.
The first call at layer 0 should use X_0 initialised via `init_state`.
"""
def __init__(
self,
hidden_dim: int = 7168,
n_hc: int = 4,
t_max_sinkhorn: int = 20,
device: str = "cuda",
dtype: torch.dtype = torch.bfloat16,
):
self.d = hidden_dim
self.n_hc = n_hc
self.K_proj = n_hc * hidden_dim # 28672 for V4-Pro
self.N_proj = n_hc + n_hc * n_hc + n_hc # 4 + 16 + 4 = 24
self.t_max = t_max_sinkhorn
self.device = device
self.dtype = dtype
# ── Learnable weights (set via load_weights) ──────────────────
# Stacked projection: b shape = (N_proj, K_proj) in FP32
# Stored as separate tensors, fused in forward if DeepGEMM available.
self.W_pre = self._buf(n_hc, self.K_proj, dtype=torch.float32) # (4, K)
self.W_res = self._buf(n_hc * n_hc, self.K_proj, dtype=torch.float32) # (16, K)
self.W_post = self._buf(n_hc, self.K_proj, dtype=torch.float32) # (4, K)
# Static biases (eq. 3-5, S^pre / S^res / S^post)
self.S_pre = self._buf(1, n_hc) # (1, 4)
self.S_res = self._buf(n_hc, n_hc) # (4, 4)
self.S_post = self._buf(n_hc, 1) # (4, 1)
# Learnable gating scalars (α), initialised small during training
# At inference these are just scalars loaded from the checkpoint.
self.alpha_pre = torch.zeros(1, device=device, dtype=torch.float32)
self.alpha_res = torch.zeros(1, device=device, dtype=torch.float32)
self.alpha_post = torch.zeros(1, device=device, dtype=torch.float32)
# Pre-allocated split buffers (set in _ensure_buffers)
self._d_split = None # (NUM_SPLITS, max_T, N_proj) FP32
self._sqr_sum_split = None # (NUM_SPLITS, max_T) FP32
self._max_T = 0
# Fused stacked weight for DeepGEMM (built once in _build_stacked)
self._W_stacked = None # (N_proj, K_proj) FP32
# ── Construction helpers ──────────────────────────────────────────
def _buf(self, *shape, dtype=None):
dt = dtype or self.dtype
return torch.empty(*shape, dtype=dt, device=self.device)
def load_weights(
self,
W_pre: torch.Tensor, # (n_hc, K) FP32
W_res: torch.Tensor, # (n_hc², K) FP32
W_post: torch.Tensor, # (n_hc, K) FP32
S_pre: torch.Tensor, # (1, n_hc)
S_res: torch.Tensor, # (n_hc, n_hc)
S_post: torch.Tensor, # (n_hc, 1)
alpha_pre: float,
alpha_res: float,
alpha_post: float,
):
"""
Load all mHC parameters from the checkpoint.
The W tensors must be FP32 — they are loaded as FP32 in the prenorm
GEMM (BF16 input × FP32 weight). Everything else can be BF16 in the
checkpoint and will be cast here.
"""
def _f32(t): return t.to(device=self.device, dtype=torch.float32).contiguous()
def _cvt(t): return t.to(device=self.device, dtype=self.dtype).contiguous()
self.W_pre = _f32(W_pre)
self.W_res = _f32(W_res)
self.W_post = _f32(W_post)
self.S_pre = _cvt(S_pre)
self.S_res = _cvt(S_res)
self.S_post = _cvt(S_post)
self.alpha_pre = torch.tensor(alpha_pre, dtype=torch.float32, device=self.device)
self.alpha_res = torch.tensor(alpha_res, dtype=torch.float32, device=self.device)
self.alpha_post = torch.tensor(alpha_post, dtype=torch.float32, device=self.device)
self._W_stacked = None # invalidate cache
def _build_stacked(self):
"""Fuse W_pre / W_res / W_post into one (N_proj, K_proj) FP32 tensor."""
self._W_stacked = torch.cat([self.W_pre, self.W_res, self.W_post], dim=0)
# Must be K-major (contiguous along K) for DeepGEMM
self._W_stacked = self._W_stacked.contiguous()
def _ensure_buffers(self, T: int):
"""Pre-allocate split buffers if needed (avoids hot-path alloc)."""
if T <= self._max_T:
return
self._d_split = torch.empty(
NUM_SPLITS, T, self.N_proj, dtype=torch.float32, device=self.device
)
self._sqr_sum_split = torch.empty(
NUM_SPLITS, T, dtype=torch.float32, device=self.device
)
self._max_T = T
# ── Forward ──────────────────────────────────────────────────────
def _project_and_rms(self, X_flat: torch.Tensor) -> torch.Tensor:
"""
Compute RMSNorm(X_flat) @ W_stacked.T → (T, N_proj) FP32.
Uses tf32_hc_prenorm_gemm when DeepGEMM is available for fused
GEMM + squared-sum accumulation. Falls back to plain BF16 matmul.
X_flat: (T, K_proj) BF16
"""
T = X_flat.shape[0]
K = self.K_proj
if _HAS_DEEP_GEMM:
if self._W_stacked is None:
self._build_stacked()
self._ensure_buffers(T)
d_s = self._d_split[:, :T, :] # view, no copy
ss_s = self._sqr_sum_split[:, :T]
# a: (T, K) BF16 b: (N, K) FP32 → d_s: (S, T, N), ss_s: (S, T)
# Both d and sqr_sum are OUTPUT tensors (written by the kernel).
deep_gemm.tf32_hc_prenorm_gemm(
X_flat.contiguous(), # a
self._W_stacked, # b (N, K) FP32
d_s, # d (S, T, N)
ss_s, # sqr_sum (S, T)
num_splits=NUM_SPLITS,
)
d_out = d_s.sum(dim=0) # (T, N)
sqr_sum = ss_s.sum(dim=0) # (T,)
else:
# Fallback: BF16 matmul + manual squared sum
if self._W_stacked is None:
self._build_stacked()
x_f32 = X_flat.float()
d_out = x_f32 @ self._W_stacked.T # (T, N)
sqr_sum = x_f32.pow(2).sum(dim=-1) # (T,)
# RMSNorm scale: multiply raw GEMM output by rsqrt(mean(x²))
# mean(x²) = sqr_sum / K → scale = sqrt(K / sqr_sum)
rms_scale = torch.sqrt(K / (sqr_sum + EPS_RMSN)) # (T,)
return (d_out * rms_scale.unsqueeze(-1)).to(self.dtype) # (T, N) in BF16
def _dynamic_params(
self, X_l: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Compute per-token A_l, B_l, C_l from the current residual state.
X_l: (T, n_hc, d)
Returns:
A_l: (T, n_hc) sigmoid-constrained input mapping
B_l: (T, n_hc, n_hc) doubly-stochastic residual transform
C_l: (T, n_hc) 2*sigmoid-constrained output mapping
"""
T, n, d = X_l.shape
assert n == self.n_hc and d == self.d
# Flatten: (T, n_hc*d)
X_flat = X_l.reshape(T, self.K_proj).to(self.dtype)
# Fused RMSNorm projection: (T, N_proj)
proj = self._project_and_rms(X_flat).float() # keep FP32 for precision
# Split into raw A / B / C
i0, i1, i2, i3 = 0, self.n_hc, self.n_hc + self.n_hc**2, self.N_proj
A_raw = proj[:, i0:i1] # (T, n_hc)
B_raw = proj[:, i1:i2] # (T, n_hc²)
C_raw = proj[:, i2:i3] # (T, n_hc)
# Add static biases and scale by learned gating factors (eq. 3-5)
S_pre = self.S_pre.float() # (1, n_hc)
S_res = self.S_res.float() # (n_hc, n_hc)
S_post = self.S_post.float() # (n_hc, 1)
A_tilde = self.alpha_pre * A_raw + S_pre # (T, n_hc)
B_tilde = self.alpha_res * B_raw + S_res.flatten().unsqueeze(0) # (T, n_hc²)
C_tilde = self.alpha_post * C_raw + S_post.flatten().unsqueeze(0) # (T, n_hc)
# Apply constraints (paper eqs. 6-8)
A_l = torch.sigmoid(A_tilde) # (T, n_hc)
C_l = 2.0 * torch.sigmoid(C_tilde) # (T, n_hc)
# B_l: exp → Sinkhorn-Knopp → doubly stochastic
B_exp = torch.exp(B_tilde).reshape(T, self.n_hc, self.n_hc)
B_l = sinkhorn_knopp(B_exp, t_max=self.t_max) # (T, n_hc, n_hc)
# Keep B_l in FP32 — the (T,4,4) bmm precision matters more than memory.
# A_l and C_l are cast to dtype for the input/output mixing multiplies.
return A_l.to(self.dtype), B_l, C_l.to(self.dtype)
# ----------------------------------------------------------------
# Public API: pre_block / post_block
# ----------------------------------------------------------------
def pre_block(
self,
X_l: torch.Tensor, # (T, n_hc, d) BF16
) -> Tuple[torch.Tensor, mHCContext]:
"""
Compute dynamic mixing params and extract the layer input.
Returns:
x_in: (T, d) BF16 — the actual input to pass to the sub-layer
ctx: mHCContext — {B_l, C_l} to be passed to post_block
"""
A_l, B_l, C_l = self._dynamic_params(X_l)
# Layer input: x_in = A_l @ X_l (per token, weighted sum of streams)
# A_l: (T, n_hc) X_l: (T, n_hc, d)
# → (T, 1, n_hc) bmm (T, n_hc, d) = (T, 1, d) → squeeze
x_in = torch.bmm(A_l.unsqueeze(1), X_l).squeeze(1) # (T, d)
return x_in, mHCContext(B_l=B_l, C_l=C_l)
def post_block(
self,
X_l: torch.Tensor, # (T, n_hc, d) BF16 — residual state BEFORE sub-layer
F_out: torch.Tensor, # (T, d) BF16 — sub-layer output
ctx: mHCContext,
) -> torch.Tensor:
"""
Apply the mHC residual update (eq. 1):
X_{l+1} = B_l @ X_l + C_l ⊗ F_out
Returns:
X_next: (T, n_hc, d) BF16
"""
# B_l is FP32, X_l is BF16 — bmm upcasts automatically in PyTorch.
BX = torch.bmm(ctx.B_l, X_l.float())
CF = ctx.C_l.unsqueeze(-1) * F_out.unsqueeze(1) # (T, n_hc, d)
return (BX + CF.float()).to(self.dtype) # (T, n_hc, d)
# ----------------------------------------------------------------
# Utility
# ----------------------------------------------------------------
@staticmethod
def init_state(
embeddings: torch.Tensor, # (T, d) BF16 — token embeddings
n_hc: int = 4,
) -> torch.Tensor:
"""
Initialise X_0 for the first layer.
The paper figure shows the embedding feeding into the first
Residual Mixing. We broadcast the embedding across all n_hc
residual streams as the simplest valid initialisation.
Returns: (T, n_hc, d) BF16
"""
return embeddings.unsqueeze(1).expand(-1, n_hc, -1).clone()
@staticmethod
def read_out(X_L: torch.Tensor) -> torch.Tensor:
"""
Extract the final hidden state from the last residual state.
Convention: stream 0 is the primary output stream (standard choice
for HC models — the first stream carries the main residual).
Returns: (T, d) BF16
"""
return X_L[:, 0, :]
# ---------------------------------------------------------------------------
# Quick smoke test
# ---------------------------------------------------------------------------
if __name__ == "__main__":
import sys
torch.manual_seed(0)
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16
D, N_HC = 7168, 4
K = N_HC * D # 28672
N_PROJ = N_HC + N_HC ** 2 + N_HC # 24
mhc = mHCLayer(hidden_dim=D, n_hc=N_HC, device=device, dtype=dtype)
# Random weights matching the expected shapes
mhc.load_weights(
W_pre = torch.randn(N_HC, K, dtype=torch.float32),
W_res = torch.randn(N_HC**2, K, dtype=torch.float32),
W_post = torch.randn(N_HC, K, dtype=torch.float32),
S_pre = torch.zeros(1, N_HC, dtype=dtype),
S_res = torch.eye(N_HC, dtype=dtype), # identity: pure residual
S_post = torch.zeros(N_HC, 1, dtype=dtype),
alpha_pre = 0.01,
alpha_res = 0.01,
alpha_post = 0.01,
)
T = 4 # 4 tokens
# ── Forward pass ────────────────────────────────────────────────
embeddings = torch.randn(T, D, dtype=dtype, device=device)
X = mHCLayer.init_state(embeddings, n_hc=N_HC)
print(f"X_0: {X.shape} (T={T}, n_hc={N_HC}, d={D})")
# Simulate a 2-layer stack
for layer_idx in range(2):
x_in, ctx = mhc.pre_block(X)
print(f"\nLayer {layer_idx}:")
print(f" x_in (to sub-layer): {x_in.shape}")
print(f" B_l: {ctx.B_l.shape}")
print(f" C_l: {ctx.C_l.shape}")
# Dummy sub-layer: identity (for testing the mHC mechanics)
F_out = x_in
X = mhc.post_block(X, F_out, ctx)
print(f" X_next: {X.shape}")
hidden = mHCLayer.read_out(X)
print(f"\nFinal hidden: {hidden.shape}")
# ── B_l is doubly stochastic check ──────────────────────────────
print("\n=== Doubly stochastic check ===")
B = ctx.B_l # (T, 4, 4) — FP32 from Sinkhorn
row_sums = B.sum(dim=-1) # (T, 4) — should all be ~1
col_sums = B.sum(dim=-2) # (T, 4) — should all be ~1
print(f" row sum range: [{row_sums.min():.6f}, {row_sums.max():.6f}] (want ≈ 1.0)")
print(f" col sum range: [{col_sums.min():.6f}, {col_sums.max():.6f}] (want ≈ 1.0)")
assert (row_sums - 1).abs().max() < 1e-3, "B_l rows do not sum to 1"
assert (col_sums - 1).abs().max() < 1e-3, "B_l cols do not sum to 1"
print(" PASSED")
# ── A_l and C_l are bounded ──────────────────────────────────────
# (Re-run dynamic params to expose A_l for checking)
A_l, B_l2, C_l = mhc._dynamic_params(X)
print(f"\n=== A_l ∈ (0,1) check ===")
print(f" A_l range: [{A_l.min():.4f}, {A_l.max():.4f}] (want ∈ (0,1))")
assert A_l.min() > 0 and A_l.max() < 1, "A_l out of sigmoid range"
print(" PASSED")
print(f"\n=== C_l ∈ (0,2) check ===")
print(f" C_l range: [{C_l.min():.4f}, {C_l.max():.4f}] (want ∈ (0,2))")
assert C_l.min() > 0 and C_l.max() < 2, "C_l out of 2*sigmoid range"
print(" PASSED")
# ── Consistency: S_res = identity → B_l ≈ doubly-stochastic I ───
print("\n=== S_res=I, alpha_res≈0 → B_l ≈ uniform matrix ===")
# With S_res = I and alpha_res ≈ 0:
# B_tilde ≈ I → exp(I) → Sinkhorn of exp(I)
# exp(I) is diag-dominant; after Sinkhorn it converges to a doubly stochastic matrix.
# We just check doubly-stochastic property is preserved (already checked above).
print(" Already verified via doubly stochastic check above.")
# ── Equivalence: T=1 decode vs T=N prefill ──────────────────────
print("\n=== Token-by-token decode == batch prefill ===")
T_big = 8
h_big = torch.randn(T_big, D, dtype=dtype, device=device)
X_batch = mHCLayer.init_state(h_big, n_hc=N_HC)
# Batch
x_in_batch, ctx_batch = mhc.pre_block(X_batch)
# Token by token
x_in_tokens = []
for t in range(T_big):
X_t = X_batch[t:t+1] # (1, n_hc, d)
x_in_t, _ = mhc.pre_block(X_t)
x_in_tokens.append(x_in_t)
x_in_seq = torch.cat(x_in_tokens, dim=0) # (T_big, d)
diff = (x_in_batch - x_in_seq).abs().max().item()
print(f" max |batch - sequential| on x_in: {diff:.6f}")
assert diff < 1e-2, f"Mismatch too large: {diff}"
print(" PASSED")
print("\nAll checks done.")
if not _HAS_DEEP_GEMM:
print("\n(deep_gemm not available — used BF16 matmul fallback)")