[Model Runner V2][Spec Decode] Use fp32 uniform threshold for acceptance (#46878)

This commit is contained in:
Giancarlo Delfin
2026-06-27 14:09:25 -07:00
committed by GitHub
parent 8bf064f8d3
commit b6caeb5a09
2 changed files with 12 additions and 5 deletions
+9 -2
View File
@@ -73,6 +73,14 @@ def tl_rand64(seed, offset, includes_zero: tl.constexpr):
return u
@triton.jit
def tl_rand32(seed, offset, includes_zero: tl.constexpr):
u = tl.rand(seed, offset)
if not includes_zero:
u = tl.maximum(u, _TL_RAND_MIN)
return u
@triton.jit
def gumbel_block_argmax(
logits,
@@ -131,8 +139,7 @@ def gumbel_block_argmax(
u = tl_rand64(gumbel_seed, block, includes_zero=False)
gumbel_noise = -tl.log(-tl.log(u))
else:
u = tl.rand(gumbel_seed, block)
u = tl.maximum(u, _TL_RAND_MIN)
u = tl_rand32(gumbel_seed, block, includes_zero=False)
# Draw the large-noise tail (which decides the argmax winner) from u -> 0,
# where fp32 has fine resolution, instead of u -> 1, where fp32 spacing is
# ~2**-24. The naive `-log(-log(u))` puts the winning tail at u -> 1,
@@ -3,7 +3,7 @@
import torch
from vllm.triton_utils import tl, tldevice, triton
from vllm.v1.worker.gpu.sample.gumbel import gumbel_block_argmax, tl_rand64
from vllm.v1.worker.gpu.sample.gumbel import gumbel_block_argmax, tl_rand32
@triton.jit
@@ -243,7 +243,7 @@ def _rejection_kernel(
if SYNTHETIC_MODE:
pos = tl.load(pos_ptr + logit_idx)
u = tl_rand64(seed, pos, includes_zero=False)
u = tl_rand32(seed, pos, includes_zero=False)
rate = tl.load(synthetic_conditional_rates_ptr + i)
# -1 is used for padded draft token ids that should be rejected.
accepted &= (u < rate) & (draft_sampled >= 0)
@@ -272,7 +272,7 @@ def _rejection_kernel(
)
target_log_prob = target_logit - target_lse
pos = tl.load(pos_ptr + logit_idx)
u = tl_rand64(seed, pos, includes_zero=False)
u = tl_rand32(seed, pos, includes_zero=False)
if HAS_DRAFT_LOGITS:
draft_logit = tl.load(
draft_logits_ptr