[Quantization] FP8 Weight Reloading for Quantized RL Rollout (#28480)

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
Kyle Sayers
2025-12-09 13:54:32 -08:00
committed by GitHub
parent 00e5cbb967
commit fccd532587
7 changed files with 206 additions and 86 deletions
@@ -27,6 +27,7 @@ from vllm.model_executor.parameter import (
ChannelQuantScaleParameter,
PerTensorScaleParameter,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.deep_gemm import (
@@ -1404,12 +1405,12 @@ def maybe_post_process_fp8_weight_block(layer: torch.nn.Module):
if should_use_deepgemm:
dg_weight, dg_weight_scale = deepgemm_post_process_fp8_weight_block(
wq=layer.weight.data,
ws=layer.weight_scale.data,
ws=layer.weight_scale_inv.data,
quant_block_shape=tuple(layer.weight_block_size),
use_e8m0=is_deep_gemm_e8m0_used(),
)
layer.weight = torch.nn.Parameter(dg_weight, requires_grad=False)
layer.weight_scale = torch.nn.Parameter(dg_weight_scale, requires_grad=False)
replace_parameter(layer, "weight", dg_weight)
replace_parameter(layer, "weight_scale_inv", dg_weight_scale)
def expert_weight_is_col_major(x: torch.Tensor) -> bool:
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
marlin_quant_input,
should_use_atomic_add_reduce,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
@@ -130,7 +131,7 @@ def prepare_fp8_layer_for_marlin(
size_n=part_size_n,
num_bits=8,
)
layer.weight = torch.nn.Parameter(marlin_qweight, requires_grad=False)
replace_parameter(layer, "weight", marlin_qweight)
# WEIGHT SCALES
# Permute scales
@@ -138,7 +139,6 @@ def prepare_fp8_layer_for_marlin(
scales = layer.weight_scale.to(layer.orig_dtype)
elif "weight_scale_inv" in dir(layer):
scales = layer.weight_scale_inv.to(layer.orig_dtype)
del layer.weight_scale_inv
group_size = -1 if weight_block_size is None else weight_block_size[1]
@@ -177,12 +177,15 @@ def prepare_fp8_layer_for_marlin(
)
if input_dtype != torch.float8_e4m3fn:
marlin_scales = fp8_fused_exponent_bias_into_scales(marlin_scales)
layer.weight_scale = torch.nn.Parameter(marlin_scales, requires_grad=False)
if hasattr(layer, "weight_scale"):
replace_parameter(layer, "weight_scale", marlin_scales)
elif hasattr(layer, "weight_scale_inv"):
replace_parameter(layer, "weight_scale_inv", marlin_scales)
if hasattr(layer, "bias") and layer.bias is not None:
assert layer.bias.shape == (part_size_n,)
bias = marlin_permute_bias(layer.bias)
layer.bias = torch.nn.Parameter(bias, requires_grad=False)
replace_parameter(layer, "bias", bias)
def prepare_moe_fp8_layer_for_marlin(
@@ -118,8 +118,11 @@ def requantize_with_max_scale(
# from disk in this case. Skip requantization in this case (since)
# we already are quantized with the single scale.
# * Sample Model: nm-testing/Phi-3-mini-128k-instruct-FP8
#
# Extra note: upon weight reloading weight_scale.ndim == 0
unfused_module_in_checkpoint = (
weight_scale[-1] > torch.finfo(torch.float8_e4m3fn).min
weight_scale.ndim != 0
and weight_scale[-1] > torch.finfo(torch.float8_e4m3fn).min
)
# If unfused checkpoint, need requanize with the single scale.