Files

459 lines
19 KiB
Python
Raw Permalink Normal View History

"""NVFP4 quantization: BF16 <-> NVFP4 conversion, scale factor computation."""
import math
import torch
import cutlass
import cutlass.cute as cute
import cutlass.torch as cutlass_torch
import cutlass.utils as utils
from dsv4.ops.layouts import ceil_div
from dsv4.kernels.gemm.grouped import (
cat_byte_reinterpretable_tensors,
stack_byte_reinterpretable_tensors,
)
E2M1_MAGNITUDES = [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0]
# Cache compiled kernels + pre-allocated workspace by cache_key
# Each entry: {'compiled': callable, 'workspace': Tensor, 'workspace_size': int}
#
# Key design decisions (Bug #1 fix):
# - cute.compile does NOT corrupt GPU memory (verified 2026-05-20 on B200).
# The original _needs_token_refill hack was a misdiagnosis. The real bug
# was elsewhere (likely OOB write or weight loading).
# - Workspace is pre-allocated per cache entry during warmup_compilation()
# and reused on subsequent calls. No torch.full() in the hot path.
# - CuTe tensor wrappers (from_dlpack + mark_layout_dynamic) are cheap
# metadata wrappers. We re-create them per call from real tensors.
# Caching them would hold stale references to tensors that get freed.
# Cached LUT for E2M1 quantization (created once per device, cudagraph-safe)
_NVFP4_STEP_LUT_CACHE = {}
def _get_step_to_idx_lut(device):
"""Get or create the E2M1 step-to-index LUT for the given device.
Cached per device to avoid CPU->CUDA copies during cudagraph capture.
Must be pre-populated during warmup (before torch.compile/cudagraph capture)
so the lock is never entered on the compiled path.
"""
# Fast path: already cached — no lock needed (torch.compile-safe)
if device in _NVFP4_STEP_LUT_CACHE:
return _NVFP4_STEP_LUT_CACHE[device]
# Slow path: first call, create the LUT
lut = torch.as_tensor(
[0, 1, 2, 3, 4, 4, 5, 5, 6, 6, 6, 7, 7],
dtype=torch.int8, device=device,
)
_NVFP4_STEP_LUT_CACHE[device] = lut
return lut
SF_VEC_SIZE = 16 # NVFP4 block size
def quantize_to_nvfp4(x_bf16, block_size=SF_VEC_SIZE):
"""Quantize BF16 tensor to NVFP4.
Args:
x_bf16: (..., D) BF16 tensor
Returns:
x_fp4: (..., D//2) float4_e2m1fn_x2 native PyTorch FP4
x_sf: (..., D//16) float8_e4m3fn block scales
global_scale: float32 scalar
"""
x_f32 = x_bf16.float()
amax = x_f32.abs().max().clamp(min=1e-8).float()
global_scale = amax / (6.0 * 448.0)
x_norm = x_f32 / global_scale
last_dim = x_norm.shape[-1]
n_blocks = ceil_div(last_dim, block_size)
if last_dim % block_size != 0:
pad_size = n_blocks * block_size - last_dim
x_norm = torch.nn.functional.pad(x_norm, (0, pad_size))
x_reshaped = x_norm.reshape(*x_norm.shape[:-1], n_blocks, block_size)
block_amax = x_reshaped.abs().amax(dim=-1)
# Detect zero blocks and underflow blocks (amax > 0 but too small for FP8).
# Smallest positive FP8 e4m3fn is 2^-9 ≈ 1.95e-3. If amax/6 < this,
# the block scale underflows to 0, and dividing x by the clamped 1e-8
# inflates values into nonzero FP4 buckets — producing wrong results.
zero_block = block_amax < (6.0 * 2.0 ** -9) # < ~0.0117
# Zero out x for zero/underflow blocks before division.
# This ensures x_scaled = 0 → FP4 nibbles = 0.
# Use scalar 0.0 instead of torch.zeros_like — no allocation, graph-safe.
x_reshaped = torch.where(zero_block.unsqueeze(-1), 0.0, x_reshaped)
block_amax = block_amax.clamp(min=1e-8)
block_scale = (block_amax / 6.0).to(torch.float8_e4m3fn)
# Force zero/underflow blocks: FP8 scale = 0 (exact zero).
block_scale = torch.where(zero_block, 0.0, block_scale)
# Nearest E2M1
block_sf_expanded = block_scale.float().unsqueeze(-1)
x_scaled = x_reshaped / block_sf_expanded.clamp(min=1e-8)
signs = torch.sign(x_scaled)
abs_scaled = x_scaled.abs().clamp(max=6.0)
half_steps = (abs_scaled * 2.0).round().clamp(0, 12).to(torch.int8)
step_to_idx = _get_step_to_idx_lut(x_bf16.device)
indices = step_to_idx[half_steps.long()]
nibbles = torch.where(signs < 0, indices + 8, indices).to(torch.uint8)
even = nibbles[..., ::2]
odd = nibbles[..., 1::2]
packed = (odd << 4) | even
packed_shape = list(x_bf16.shape)
packed_shape[-1] = last_dim // 2
x_fp4 = packed.view(torch.float4_e2m1fn_x2).reshape(packed_shape)
sf_shape = list(x_bf16.shape[:-1]) + [n_blocks]
block_scale = block_scale.reshape(sf_shape)
return x_fp4, block_scale, global_scale
def quantize_activation_nvfp4(x_bf16, global_scale, block_size=SF_VEC_SIZE):
"""Quantize BF16 activation tensor to NVFP4 (cudagraph-safe).
Unlike quantize_to_nvfp4(), this takes a pre-computed global_scale
instead of computing it via .max() (which forces CPU-GPU sync).
All operations are pure GPU with no CPU-GPU syncs.
Args:
x_bf16: (..., D) BF16 tensor
global_scale: float32 scalar (pre-computed, NOT from .max())
block_size: NVFP4 block size
Returns:
x_fp4: (..., D//2) float4_e2m1fn_x2
x_sf: (..., D//16) float8_e4m3fn
"""
x_f32 = x_bf16.float()
x_norm = x_f32 / global_scale
last_dim = x_norm.shape[-1]
n_blocks = ceil_div(last_dim, block_size)
if last_dim % block_size != 0:
pad_size = n_blocks * block_size - last_dim
x_norm = torch.nn.functional.pad(x_norm, (0, pad_size))
x_reshaped = x_norm.reshape(*x_norm.shape[:-1], n_blocks, block_size)
block_amax = x_reshaped.abs().amax(dim=-1)
# Detect zero blocks and underflow blocks (same threshold as quantize_to_nvfp4).
zero_block = block_amax < (6.0 * 2.0 ** -9)
x_reshaped = torch.where(zero_block.unsqueeze(-1), 0.0, x_reshaped)
block_amax = block_amax.clamp(min=1e-8, max=6.0 * 448.0) # E4M3 max = 448
block_scale = (block_amax / 6.0).to(torch.float8_e4m3fn)
block_scale = torch.where(zero_block, 0.0, block_scale)
block_sf_expanded = block_scale.float().unsqueeze(-1)
x_scaled = x_reshaped / block_sf_expanded.clamp(min=1e-8)
signs = torch.sign(x_scaled)
abs_scaled = x_scaled.abs().clamp(max=6.0)
half_steps = (abs_scaled * 2.0).round().clamp(0, 12).to(torch.int8)
step_to_idx = _get_step_to_idx_lut(x_bf16.device)
indices = step_to_idx[half_steps.long()]
nibbles = torch.where(signs < 0, indices + 8, indices).to(torch.uint8)
even = nibbles[..., ::2]
odd = nibbles[..., 1::2]
packed = (odd << 4) | even
packed_shape = list(x_bf16.shape)
packed_shape[-1] = last_dim // 2
x_fp4 = packed.view(torch.float4_e2m1fn_x2).reshape(packed_shape)
sf_shape = list(x_bf16.shape[:-1]) + [n_blocks]
block_scale = block_scale.reshape(sf_shape)
return x_fp4, block_scale
def quantize_weight_to_nvfp4(w_bf16, block_size=SF_VEC_SIZE):
"""Quantize BF16 weight matrix to NVFP4.
The weight is (K, N) where K is the input dim (packed dimension).
Block scales are computed along K (dim 0).
Args:
w_bf16: (K, N) BF16 weight matrix
Returns:
w_fp4: (K//2, N) float4_e2m1fn_x2 K is the packed dim
w_sf: (K//16, N) float8_e4m3fn block scales along K
global_scale: float32 scalar
"""
K, N = w_bf16.shape
w_f32 = w_bf16.float()
amax = w_f32.abs().max().clamp(min=1e-8).float()
global_scale = amax / (6.0 * 448.0)
w_norm = w_f32 / global_scale
k_blocks = ceil_div(K, block_size)
if K % block_size != 0:
w_norm = torch.nn.functional.pad(w_norm, (0, 0, 0, k_blocks * block_size - K))
w_reshaped = w_norm.reshape(k_blocks, block_size, N)
w_block_amax = w_reshaped.abs().amax(dim=1)
# Detect zero blocks and underflow blocks (same threshold).
zero_block = w_block_amax < (6.0 * 2.0 ** -9)
w_reshaped = torch.where(zero_block.unsqueeze(1),
torch.zeros_like(w_reshaped), w_reshaped)
w_block_amax = w_block_amax.clamp(min=1e-8)
w_sf = (w_block_amax / 6.0).to(torch.float8_e4m3fn)
w_sf = torch.where(zero_block, torch.zeros_like(w_sf), w_sf)
w_block_sf = w_sf.float().unsqueeze(1)
w_scaled = w_reshaped / w_block_sf.clamp(min=1e-8)
signs = torch.sign(w_scaled)
abs_scaled = w_scaled.abs().clamp(max=6.0)
half_steps = (abs_scaled * 2.0).round().clamp(0, 12).to(torch.int8)
step_to_idx = _get_step_to_idx_lut(w_bf16.device)
indices = step_to_idx[half_steps.long()]
nibbles = torch.where(signs < 0, indices + 8, indices).to(torch.uint8)
even = nibbles[:, ::2, :]
odd = nibbles[:, 1::2, :]
packed = (odd << 4) | even
w_fp4 = packed.reshape(K // 2, N).view(torch.float4_e2m1fn_x2)
return w_fp4, w_sf, global_scale
# ── Scale Factor Assembly ─────────────────────────────────────────────
def deinterleave_quantize_nvfp4_cuda(fused_bf16, intermediate, global_scale, granularity=8):
"""De-interleave + quantize fused SwiGLU output using a custom CUDA kernel.
Single kernel launch, no Python loop. 4x faster than the Python path.
Args:
fused_bf16: (M, 2*intermediate) BF16 fused L1 output with interleaved gate/up
intermediate: intermediate dimension (e.g., 3072)
global_scale: pre-computed global scale for quantization
granularity: interleave granularity in BF16 columns (default 8)
Returns:
x_fp4: (M, intermediate//2) float4_e2m1fn_x2 quantized SwiGLU
x_sf: (M, intermediate//16) float8_e4m3fn block scales
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("deinterleave_quantize_nvfp4", ["deinterleave_quantize.cu"])
return mod.deinterleave_quantize_nvfp4(fused_bf16, intermediate, granularity, global_scale)
def deinterleave_amax_quantize_nvfp4_fused(fused_bf16, intermediate, divisor=6.0 * 448.0, granularity=8):
"""Fused deinterleave + amax + quantize: zero CPU syncs, two kernel launches.
For the MoE fused_swiglu L2 path. Two-kernel approach (correct):
Kernel 1: compute_amax_gsa on the de-interleaved values (GPU-only)
Kernel 2: deinterleave_quantize_from_buffer using gsa from GPU buffer
Args:
fused_bf16: (M, 2*intermediate) BF16 fused L1 output
intermediate: intermediate dimension
divisor: gsa = amax / divisor. Default 2688.0.
granularity: interleave granularity (default 8)
Returns:
x_fp4: (M, intermediate//2) float4_e2m1fn_x2
x_sf: (M, intermediate//16) float8_e4m3fn
gsa: (M,) float32 GPU tensor per-row global scale for L2 GEMM
"""
from dsv4.kernels.cuda.loader import get_cuda_module
# Compute gsa from the fused output
amax_mod = get_cuda_module("amax_gsa", ["amax_gsa.cu"])
gsa_gpu = amax_mod.compute_amax_gsa(fused_bf16, divisor)
M = fused_bf16.shape[0]
if gsa_gpu.dim() == 0:
gsa_gpu = gsa_gpu.reshape(1).expand(M).contiguous()
elif gsa_gpu.shape[0] == 1 and M > 1:
gsa_gpu = gsa_gpu.expand(M).contiguous()
# Deinterleave + quantize using gsa from GPU buffer
quant_mod = get_cuda_module("fused_amax_quantize", ["fused_amax_quantize.cu"])
x_fp4, x_sf = quant_mod.deinterleave_quantize_from_buffer(fused_bf16, intermediate, granularity, gsa_gpu)
return x_fp4, x_sf, gsa_gpu
def compute_amax_gsa_gpu(x_bf16, divisor=6.0 * 448.0):
"""Compute gsa = max(|x|) / divisor on GPU. No CPU sync.
Returns a scalar GPU tensor (not a Python float!).
NOTE: Prefer quantize_nvfp4_gpu_fused() which does amax+quantize in
one kernel launch. This function is kept for cases where you need gsa
without quantization.
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("amax_gsa", ["amax_gsa.cu"])
return mod.compute_amax_gsa(x_bf16, divisor)
def quantize_nvfp4_gpu_fused(x_bf16, divisor=6.0 * 448.0):
"""Fused amax + gsa + quantize: zero CPU syncs, two kernel launches.
Two-kernel approach (correct cross-CTA reduction):
Kernel 1: compute_amax_gsa row-wise amax gsa on GPU (no .item())
Kernel 2: quantize_nvfp4_from_buffer quantize using gsa from GPU buffer
The previous single-kernel approach had a race condition: the cross-CTA
shared memory reduction used __syncthreads() which only syncs within a
CTA, not across CTAs in the same grid. CTA 0 could read s_amax[b] before
CTA b had written it, producing garbage gsa values.
Args:
x_bf16: (M, N) BF16 tensor. N must be a multiple of 16.
divisor: gsa = amax / divisor. Default 6.0 * 448.0 = 2688.0.
Returns:
x_fp4: (M, N//2) float4_e2m1fn_x2
x_sf: (M, N//16) float8_e4m3fn
gsa: (M,) float32 GPU tensor per-row global scale for GEMM
"""
# CUDA kernels require contiguous input — column slices from deinterleave are non-contiguous.
# For CUDA graph capture, this MUST be contiguous at graph construction time.
# The .contiguous() call is a no-op when already contiguous (no allocation).
if not x_bf16.is_contiguous():
x_bf16 = x_bf16.contiguous()
from dsv4.kernels.cuda.loader import get_cuda_module
amax_mod = get_cuda_module("amax_gsa", ["amax_gsa.cu"])
gsa_gpu = amax_mod.compute_amax_gsa(x_bf16, divisor) # scalar GPU tensor
# Broadcast to (M,) for the quantize-from-buffer kernel.
# CUDA-graph-safe approach:
# - For M=1 decode (graph-captured): just reshape to (1,) — no allocation.
# - For M>1 prefill (not graph-captured): expand + contiguous is fine.
M = x_bf16.shape[0]
if gsa_gpu.dim() == 0:
gsa_gpu = gsa_gpu.reshape(1) # scalar → (1,) — no allocation
if M > 1:
gsa_gpu = gsa_gpu.expand(M).contiguous() # (M,) — allocation OK for prefill
# For M=1: gsa_gpu is (1,) contiguous — zero allocation
quant_mod = get_cuda_module("fused_amax_quantize", ["fused_amax_quantize.cu"])
x_fp4, x_sf = quant_mod.quantize_nvfp4_from_buffer(x_bf16, gsa_gpu)
return x_fp4, x_sf, gsa_gpu
def quantize_nvfp4_gpu(x_bf16, global_scale):
"""Quantize BF16 tensor to NVFP4 using a custom CUDA kernel (GPU-only, no CPU sync).
Replaces quantize_activation_nvfp4() which uses .amax() (CPU sync).
The global_scale must be pre-computed (from warmup or known value).
NOTE: Prefer quantize_nvfp4_gpu_fused() which also computes gsa on GPU.
This function is kept for cases where global_scale is already known.
Args:
x_bf16: (M, N) BF16 tensor. N must be a multiple of 16.
global_scale: float32 scalar (pre-computed, NOT from .max())
Returns:
x_fp4: (M, N//2) float4_e2m1fn_x2
x_sf: (M, N//16) float8_e4m3fn
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("quantize_nvfp4", ["quantize_nvfp4.cu"])
return mod.quantize_nvfp4(x_bf16, global_scale)
class QuantizedActivation:
"""Pre-quantized NVFP4 activation tensor.
Carries the FP4 data, block scales, and per-row global scale
so downstream Nvfp4Linear calls can skip quantization and go
straight to GEMM.
Created by rmsnorm_quantize_nvfp4() or quantize_nvfp4_gpu_fused().
Consumed by Nvfp4Linear.run_from_quantized().
"""
__slots__ = ['x_fp4', 'x_sf', 'gsa', 'inv_rms', 'num_tokens']
def __init__(self, x_fp4, x_sf, gsa, inv_rms=None):
self.x_fp4 = x_fp4 # (M, N//2) FP4
self.x_sf = x_sf # (M, N//16) E4M3
self.gsa = gsa # (M,) FP32
self.inv_rms = inv_rms # (M,) FP32, optional
self.num_tokens = x_fp4.shape[0]
def dequantize_nvfp4(x_fp4, x_sf, gsa, shape=None):
"""Dequantize NVFP4 → BF16 using the CUDA dequant kernel.
Args:
x_fp4: (M, N//2) FP4 packed
x_sf: (M, N//16) E4M3 block scales
gsa: (M,) or (M, 1) or (1,) FP32 global scale per row
shape: unused, kept for API compat
Returns:
(M, N) BF16 tensor
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("dequant_nvfp4", ["dequant_nvfp4.cu"])
if gsa.dim() == 2:
gsa = gsa.squeeze(1) # (M, 1) → (M,)
# dequant kernel expects uint8 for both fp4 and sf
if x_fp4.dtype != torch.uint8:
x_fp4 = x_fp4.view(torch.uint8)
if x_sf.dtype != torch.uint8:
x_sf = x_sf.view(torch.uint8)
return mod.dequant_nvfp4(x_fp4, x_sf, gsa)
def mhc_rmsnorm_quantize_nvfp4(X_l, A_l, norm_weight, eps=1e-6, divisor=6.0 * 448.0):
"""Fused mHC pre_block + RMSNorm + NVFP4 quantize: 2 kernel launches total.
Replaces: bmm (1 launch) + rmsnorm (4+ launches) + quantize (2 launches)
Total unfused: 7+ launches per site × 122 sites = 854+ launches/token
Fused: 2 launches per site × 122 sites = 244 launches 610 launches saved/token.
Args:
X_l: (M, n_hc, N) BF16 tensor. n_hc must be <= 4, N multiple of 16.
A_l: (M, n_hc) BF16 tensor. Softmax weights from mHC._dynamic_params.
norm_weight: (N,) FP32 RMSNorm weight.
eps: RMSNorm epsilon (default 1e-6).
divisor: gsa = amax / divisor. Default 6.0 * 448.0 = 2688.0.
Returns:
QuantizedActivation with x_fp4, x_sf, gsa, inv_rms
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("fused_mhc_rmsnorm_quantize", ["fused_mhc_rmsnorm_quantize.cu"])
x_fp4, x_sf, gsa, inv_rms = mod.mhc_rmsnorm_quantize_nvfp4(X_l, A_l, norm_weight, eps, divisor)
return QuantizedActivation(x_fp4, x_sf, gsa, inv_rms)
def rmsnorm_quantize_nvfp4(x_bf16, norm_weight, eps=1e-6, divisor=6.0 * 448.0):
"""Fused RMSNorm + amax + NVFP4 quantize: 2 kernel launches total.
Replaces the unfused path:
rmsnorm(x, weight) 4+ BF16 launches
quantize_nvfp4_gpu_fused(rmsnormed) 2 kernel launches + amax
Total unfused: 6+ launches per call × 122 calls/layer-step = 732+ launches/token
Fused: 2 kernel launches per call × 122 calls = 244 launches 488 launches saved/token.
Two-kernel approach (correct cross-CTA reduction):
Kernel 1: compute RMS + amax of normalized output gsa per row (GPU buffer)
Kernel 2: normalize + quantize using gsa from GPU buffer (no CPU sync)
Args:
x_bf16: (M, N) BF16 tensor. N must be a multiple of 16.
norm_weight: (N,) FP32 RMSNorm weight.
eps: RMSNorm epsilon (default 1e-6).
divisor: gsa = amax / divisor. Default 6.0 * 448.0 = 2688.0.
Returns:
x_fp4: (M, N//2) FP4 packed (uint8 view of float4_e2m1fn_x2)
x_sf: (M, N//16) E4M3 block scales
gsa: (M,) FP32 per-row global scale for GEMM
inv_rms: (M,) FP32 per-row 1/RMS (useful for downstream if needed)
"""
from dsv4.kernels.cuda.loader import get_cuda_module
mod = get_cuda_module("fused_rmsnorm_quantize", ["fused_rmsnorm_quantize.cu"])
x_fp4, x_sf, gsa, inv_rms = mod.rmsnorm_quantize_nvfp4(x_bf16, norm_weight, eps, divisor)
return QuantizedActivation(x_fp4, x_sf, gsa, inv_rms)