[watermarking] add context deduplication support to speculative decoding (#56807)
Signed-off-by: Raphael Rialland <raphael.rialland@mistral.ai> Signed-off-by: Simon Veitner <sveitner@redhat.com> Co-authored-by: Simon Veitner <sveitner@redhat.com> Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
This commit is contained in:
co-authored by
Simon Veitner
Lucas Wilkinson
parent
73a5831127
commit
0f8b398158
@@ -135,7 +135,12 @@ class DSparkSpeculator(DFlashSpeculator):
|
||||
|
||||
# sample_pos is the predicted token's position P. Sampling keys a draw
|
||||
# by the position before the sampled token, P-1.
|
||||
sampled = gumbel_sample(
|
||||
sampler = (
|
||||
gumbel_sample
|
||||
if self.draft_watermarker is None
|
||||
else self.draft_watermarker.sample
|
||||
)
|
||||
return sampler(
|
||||
logits,
|
||||
idx_map,
|
||||
self.temperature,
|
||||
@@ -147,11 +152,6 @@ class DSparkSpeculator(DFlashSpeculator):
|
||||
logits_cache_col=self._step_cols[step],
|
||||
use_fp64=self.use_fp64_gumbel,
|
||||
)
|
||||
if self.draft_watermarker is not None:
|
||||
sampled = self.draft_watermarker.sample(
|
||||
logits, sampled, idx_map, self.temperature
|
||||
)
|
||||
return sampled
|
||||
|
||||
def _sample_sequential(self, num_reqs: int, head_hidden: torch.Tensor) -> None:
|
||||
if self._draft_topk is not None:
|
||||
|
||||
@@ -226,11 +226,18 @@ class RejectionSampler:
|
||||
if self.watermark_key is None:
|
||||
return {}
|
||||
assert isinstance(self.sampler, GPUWatermarkSampler)
|
||||
contexts = self.sampler._get_contexts(
|
||||
expanded_idx_mapping, expanded_local_pos, draft_sampled
|
||||
)
|
||||
watermarking_skip_mask = None
|
||||
if self.sampler.deduplicate_contexts != "none":
|
||||
watermarking_skip_mask = self.sampler._get_repeated_contexts(
|
||||
expanded_idx_mapping, contexts, expanded_local_pos
|
||||
)
|
||||
return {
|
||||
"contexts": self.sampler._get_contexts(
|
||||
expanded_idx_mapping, expanded_local_pos, draft_sampled
|
||||
),
|
||||
"contexts": contexts,
|
||||
"watermarking": self.sampler.watermarking.gpu,
|
||||
"watermarking_skip_mask": watermarking_skip_mask,
|
||||
"watermark_key": self.watermark_key,
|
||||
}
|
||||
|
||||
|
||||
@@ -564,7 +564,6 @@ def _rejection_kernel(
|
||||
# stores the target argmax upon first rejection, so it rejects the
|
||||
# placeholder via `accepted` instead.
|
||||
verifying &= is_valid_draft
|
||||
|
||||
if verifying:
|
||||
pos = tl.load(pos_ptr + logit_idx)
|
||||
u = tl_rand32(seed, pos, includes_zero=False)
|
||||
@@ -719,6 +718,8 @@ def _seeded_resample_argmax(
|
||||
0, # logits_cache_stride_0
|
||||
0, # logits_cache_stride_1
|
||||
None, # logits_cache_col_ptr
|
||||
None, # logits_cache_source_ptr
|
||||
0, # logits_cache_source_stride
|
||||
vocab_size,
|
||||
IS_DRAFTING=False,
|
||||
APPLY_TEMPERATURE=False,
|
||||
@@ -766,6 +767,7 @@ def _resample_kernel(
|
||||
contexts_stride,
|
||||
# [max_num_reqs], uint8 view of a bool tensor
|
||||
watermarking_ptr,
|
||||
watermarking_skip_mask_ptr,
|
||||
watermark_key_0,
|
||||
watermark_key_1,
|
||||
vocab_size,
|
||||
@@ -775,6 +777,7 @@ def _resample_kernel(
|
||||
USE_BLOCK_VERIFICATION: tl.constexpr,
|
||||
CONTEXT_WIDTH: tl.constexpr,
|
||||
WATERMARK: tl.constexpr,
|
||||
DEDUPLICATE_CONTEXTS: tl.constexpr,
|
||||
):
|
||||
req_idx = tl.program_id(0)
|
||||
resample_idx = tl.load(rejected_step_ptr + req_idx)
|
||||
@@ -807,6 +810,25 @@ def _resample_kernel(
|
||||
).to(tl.float32)
|
||||
|
||||
# Compute the residual logits to resample the rejected token from.
|
||||
is_watermarked = False
|
||||
if WATERMARK:
|
||||
is_watermarked = (
|
||||
tl.load(
|
||||
watermarking_ptr + req_state_idx,
|
||||
mask=req_state_idx >= 0,
|
||||
other=0,
|
||||
)
|
||||
!= 0
|
||||
)
|
||||
|
||||
skip_watermarking = False
|
||||
if DEDUPLICATE_CONTEXTS:
|
||||
skip_watermarking = (
|
||||
(tl.load(watermarking_skip_mask_ptr + resample_token_idx) != 0)
|
||||
& is_watermarked
|
||||
& (temp != 0.0)
|
||||
)
|
||||
|
||||
if is_bonus or not is_valid_rejected_draft:
|
||||
# Bonus token (no rejections) or -1 placeholder token. In either case,
|
||||
# directly use the target logits.
|
||||
@@ -868,15 +890,7 @@ def _resample_kernel(
|
||||
# Resample the rejected/bonus token.
|
||||
if WATERMARK:
|
||||
# Padded and greedy rows retain the stock draw.
|
||||
is_watermarked = (
|
||||
tl.load(
|
||||
watermarking_ptr + req_state_idx,
|
||||
mask=req_state_idx >= 0,
|
||||
other=0,
|
||||
)
|
||||
!= 0
|
||||
)
|
||||
if is_watermarked & (temp != 0.0):
|
||||
if is_watermarked & (temp != 0.0) & ~skip_watermarking:
|
||||
watermark_value, idx = philox_gumbel_block_argmax(
|
||||
residual_logits,
|
||||
mask,
|
||||
@@ -1023,6 +1037,7 @@ def rejection_sample(
|
||||
use_block_verification: bool = False,
|
||||
contexts: torch.Tensor | None = None,
|
||||
watermarking: torch.Tensor | None = None,
|
||||
watermarking_skip_mask: torch.Tensor | None = None,
|
||||
watermark_key: int | None = None,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
assert target_logits.ndim == 2 and target_logits.stride(-1) == 1
|
||||
@@ -1036,9 +1051,13 @@ def rejection_sample(
|
||||
assert watermark == (watermarking is not None) == (watermark_key is not None), (
|
||||
"contexts, watermarking and watermark_key must be set together."
|
||||
)
|
||||
assert watermarking_skip_mask is None or watermark, (
|
||||
"watermarking_skip_mask requires watermarking."
|
||||
)
|
||||
contexts_stride = 0
|
||||
context_width = 1
|
||||
watermarking_bytes: torch.Tensor | None = None
|
||||
watermarking_skip_mask_bytes: torch.Tensor | None = None
|
||||
watermark_key_0 = 0
|
||||
watermark_key_1 = 0
|
||||
if contexts is not None:
|
||||
@@ -1053,6 +1072,12 @@ def rejection_sample(
|
||||
contexts_stride = contexts.stride(0)
|
||||
context_width = contexts.shape[-1]
|
||||
watermarking_bytes = watermarking.view(torch.uint8)
|
||||
if watermarking_skip_mask is not None:
|
||||
assert watermarking_skip_mask.shape == (num_logits,)
|
||||
assert watermarking_skip_mask.dtype == torch.bool
|
||||
if not watermarking_skip_mask.is_contiguous():
|
||||
watermarking_skip_mask = watermarking_skip_mask.contiguous()
|
||||
watermarking_skip_mask_bytes = watermarking_skip_mask.view(torch.uint8)
|
||||
watermark_key_0 = watermark_key & 0xFFFFFFFF
|
||||
watermark_key_1 = watermark_key >> 32
|
||||
draft_logits_stride_0 = 0
|
||||
@@ -1269,6 +1294,7 @@ def rejection_sample(
|
||||
contexts,
|
||||
contexts_stride,
|
||||
watermarking_bytes,
|
||||
watermarking_skip_mask_bytes,
|
||||
watermark_key_0,
|
||||
watermark_key_1,
|
||||
vocab_size,
|
||||
@@ -1278,6 +1304,7 @@ def rejection_sample(
|
||||
USE_BLOCK_VERIFICATION=use_block_verification,
|
||||
CONTEXT_WIDTH=context_width,
|
||||
WATERMARK=watermark,
|
||||
DEDUPLICATE_CONTEXTS=watermarking_skip_mask is not None,
|
||||
)
|
||||
|
||||
# Insert the resampled tokens into the output sampled.
|
||||
|
||||
@@ -38,6 +38,7 @@ from vllm.v1.worker.gpu.spec_decode.acceptance_estimator import (
|
||||
from vllm.v1.worker.utils import AttentionGroup
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.v1.watermarking.gpu_sampler import GPUWatermarkSampler
|
||||
from vllm.v1.worker.gpu.pcp_manager import PCPManager
|
||||
|
||||
logger = init_logger(__name__)
|
||||
@@ -198,6 +199,9 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
self.max_num_reqs,
|
||||
device,
|
||||
watermark_config.allow_target_only_watermarking,
|
||||
self.num_speculative_steps,
|
||||
watermark_config.deduplicate_contexts,
|
||||
watermark_config.deduplicate_contexts_max_history,
|
||||
)
|
||||
|
||||
self.supports_mm_inputs = False
|
||||
@@ -411,7 +415,12 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
) -> torch.Tensor:
|
||||
if draft_logits is not None:
|
||||
logits = self.model.compute_logits(hidden_states)
|
||||
sampled = gumbel_sample(
|
||||
sampler = (
|
||||
gumbel_sample
|
||||
if self.draft_watermarker is None
|
||||
else self.draft_watermarker.sample
|
||||
)
|
||||
sampled = sampler(
|
||||
logits,
|
||||
idx_mapping,
|
||||
temperature,
|
||||
@@ -423,10 +432,6 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
logits_cache_col=draft_step,
|
||||
use_fp64=self.use_fp64_gumbel,
|
||||
)
|
||||
if self.draft_watermarker is not None:
|
||||
sampled = self.draft_watermarker.sample(
|
||||
logits, sampled, idx_mapping, temperature
|
||||
)
|
||||
elif self.use_local_argmax_reduction:
|
||||
return self.model.get_top_tokens(hidden_states)
|
||||
else:
|
||||
@@ -462,11 +467,19 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
self.acceptance_estimator.step(idx_mapping, num_sampled, num_rejected)
|
||||
|
||||
def prepare_watermarking(
|
||||
self, contexts: torch.Tensor, watermarking: torch.Tensor
|
||||
self,
|
||||
sampler: "GPUWatermarkSampler",
|
||||
idx_mapping: torch.Tensor,
|
||||
) -> None:
|
||||
if self.draft_watermarker is None:
|
||||
return
|
||||
self.draft_watermarker.prepare(contexts, watermarking)
|
||||
self.draft_watermarker.prepare(
|
||||
sampler._get_contexts(idx_mapping),
|
||||
sampler.watermarking.gpu[idx_mapping],
|
||||
sampler.req_states.all_token_ids.gpu,
|
||||
sampler.req_states.prompt_len.gpu,
|
||||
sampler.req_states.total_len.gpu,
|
||||
)
|
||||
|
||||
def _copy_request_inputs(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user