apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 21 Packages Projects Releases Wiki Activity

[BugFix][Speculative Decoding] Correct index calculation in speculate decoding operators (#7121)

Browse Source 
- Fix accept_idx calculation in spec_set_value_by_stop_seqs
- Fix condition check from < to <= for token matching
- Fix accept_tokens indexing logic
- Remove unnecessary -1 in current_step comparison for max_think_len

Co-authored-by: guanshihui] <guanshihui@baidu.com>
This commit is contained in:
lonelygsh
2026-04-01 20:36:53 +08:00
committed by GitHub
parent 7a2e33098f
commit ba1aa1edff
3 changed files with 34 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
custom_ops/gpu_ops/speculate_decoding/speculate_limit_thinking_content_length.cu
+1 -4
View File
@@ -98,10 +98,7 @@ __global__ void speculate_limit_thinking_content_length_kernel(
if (max_think_len > 0) { if (max_think_len > 0) {
// A) 超长触发：到达 max_think_len 时开始注入（从本 token 起输出 // A) 超长触发：到达 max_think_len 时开始注入（从本 token 起输出
// inject_token_ids[0]） // inject_token_ids[0]）
if (status == 0 && if (status == 0 && current_step == max_think_len) {
(current_step - 1) ==
max_think_len) { // current_step - 1 是因为 speculate_verify 里
// step_idx + 1 了
status = (inject_len > 0) ? 1 : done_status; status = (inject_len > 0) ? 1 : done_status;
} }
} else if (max_think_len == 0) { } else if (max_think_len == 0) {
custom_ops/gpu_ops/speculate_decoding/speculate_set_stop_value_multi_seqs.cu
+4 -4
View File
@@ -58,7 +58,7 @@ __global__ void spec_set_value_by_stop_seqs(bool *stop_flags,
bool is_end = false; bool is_end = false;
// 遍历起始位置 // 遍历起始位置
for (; accept_idx <= accept_num - 1 && !is_end; accept_idx++) { for (; accept_idx <= accept_num - 1 && !is_end; accept_idx++) {
if (step_idx_now + accept_idx + 1 < stop_seq_len) { if (step_idx_now - accept_num + accept_idx + 1 < stop_seq_len) {
#ifdef DEBUG_SPEC_STOP_SEQS #ifdef DEBUG_SPEC_STOP_SEQS
printf("num %d < stop_seq_len %d\n", printf("num %d < stop_seq_len %d\n",
step_idx_now - accept_num + accept_idx + 1, step_idx_now - accept_num + accept_idx + 1,
@@ -71,7 +71,7 @@ __global__ void spec_set_value_by_stop_seqs(bool *stop_flags,
int64_t cur_token_idx = -1; int64_t cur_token_idx = -1;
// 通过当前值判断 token 是在 pre_ids 还是 accept_token 里 // 通过当前值判断 token 是在 pre_ids 还是 accept_token 里
if (stop_seq_len - 1 - i < accept_idx) { if (stop_seq_len - 1 - i <= accept_idx) {
#ifdef DEBUG_SPEC_STOP_SEQS #ifdef DEBUG_SPEC_STOP_SEQS
printf( printf(
"AcceptTokens bid:%d. tid:%d, accept_idx:%d, " "AcceptTokens bid:%d. tid:%d, accept_idx:%d, "
@@ -83,7 +83,7 @@ __global__ void spec_set_value_by_stop_seqs(bool *stop_flags,
accept_idx - (stop_seq_len - 1 - i) - 1); accept_idx - (stop_seq_len - 1 - i) - 1);
#endif #endif
cur_token_idx = cur_token_idx =
accept_tokens_now[accept_idx - (stop_seq_len - 1 - i) - 1]; accept_tokens_now[accept_idx - (stop_seq_len - 1 - i)];
} else { } else {
#ifdef DEBUG_SPEC_STOP_SEQS #ifdef DEBUG_SPEC_STOP_SEQS
printf( printf(
@@ -98,7 +98,7 @@ __global__ void spec_set_value_by_stop_seqs(bool *stop_flags,
(stop_seq_len - 1 - i)); (stop_seq_len - 1 - i));
#endif #endif
int pre_ids_idx = int pre_ids_idx =
step_idx_now + accept_idx - (stop_seq_len - 1 - i); step_idx_now - accept_num + accept_idx - (stop_seq_len - 1 - i);
// EC3 // EC3
// 特殊拼接会导致input_ids最后一位无特殊token，即pre_ids[0]可能为23, // 特殊拼接会导致input_ids最后一位无特殊token，即pre_ids[0]可能为23,
// 导致异常结束 // 导致异常结束
tests/operators/test_speculate_set_stop_value_multi_seqs.py
+29 -27
View File
@@ -175,17 +175,19 @@ def reference_spec_set_stop_value_multi_seqs(inputs: Dict[str, Any]) -> Dict[str
accept_idx = 0 accept_idx = 0
is_end = False is_end = False
while accept_idx <= an - 1 and not is_end: while accept_idx <= an - 1 and not is_end:
if step_idx_now + accept_idx + 1 < stop_seq_len: if step_idx_now - an + accept_idx + 1 < stop_seq_len:
accept_idx += 1 accept_idx += 1
continue continue
# Check one stop_seq match # Check one stop_seq match
for i in range(stop_seq_len - 1, -1, -1): for i in range(stop_seq_len - 1, -1, -1):
cur_token_idx = -1 cur_token_idx = -1
if stop_seq_len - 1 - i < accept_idx: # 注意:新版本kernel改成了 <=,并且去掉了 -1
cur_token_idx = accept_tokens_now[accept_idx - (stop_seq_len - 1 - i) - 1] if stop_seq_len - 1 - i <= accept_idx:
cur_token_idx = accept_tokens_now[accept_idx - (stop_seq_len - 1 - i)]
else: else:
pre_ids_idx = step_idx_now + accept_idx - (stop_seq_len - 1 - i) # 新版本:step_idx已经包含accept_num,所以要减去
pre_ids_idx = step_idx_now - an + accept_idx - (stop_seq_len - 1 - i)
if pre_ids_idx <= 0: if pre_ids_idx <= 0:
break break
cur_token_idx = pre_ids_now[pre_ids_idx] cur_token_idx = pre_ids_now[pre_ids_idx]
@@ -290,12 +292,12 @@ class TestSpeculateSetStopValueMultiSeqs(unittest.TestCase):
inputs["prompt_lens"][:] = 0 inputs["prompt_lens"][:] = 0
inputs["step_idx"][:] = 6 inputs["step_idx"][:] = 6
inputs["accept_num"][:] = 3 inputs["accept_num"][:] = 3
# Kernel matching at accept_idx=2 (3rd token, 0-indexed): # stop_seq spans pre_ids and accept_tokens
# i=2(last): stop_seq_len-1-i=0 < accept_idx(2) -> accept_tokens[2-0-1]=accept_tokens[1] # For accept_idx=1: step_idx_now - accept_num + 1 + 1 = 6-3+1+1 = 5 >= stop_seq_len=3, so we check
# i=1: stop_seq_len-1-i=1 < accept_idx(2) -> accept_tokens[2-1-1]=accept_tokens[0] # i=2: stop_seq_len-1-i=0 <= accept_idx(1) -> accept_tokens[1-0] = accept_tokens[1] = 22
# i=0: stop_seq_len-1-i=2 >= accept_idx(2) -> pre_ids[step_idx+2-(3-1-0)]=pre_ids[6] # i=1: stop_seq_len-1-i=1 <= accept_idx(1) -> accept_tokens[1-1] = accept_tokens[0] = 11
# So stop_seq should be [pre_ids[6], accept_tokens[0], accept_tokens[1]] # i=0: stop_seq_len-1-i=2 > accept_idx(1) -> pre_ids_idx = 6-3+1-(3-1-0) = 4-2 = 2 -> pre_ids[2] = 99
inputs["token_ids_all"][0, 6] = 99 inputs["token_ids_all"][0, 2] = 99
inputs["accept_tokens"][0, :3] = [11, 22, 33] inputs["accept_tokens"][0, :3] = [11, 22, 33]
inputs["stop_seqs"][0, 0, :3] = [99, 11, 22] inputs["stop_seqs"][0, 0, :3] = [99, 11, 22]
inputs["stop_seqs_len"][0, 0] = 3 inputs["stop_seqs_len"][0, 0] = 3
@@ -303,9 +305,9 @@ class TestSpeculateSetStopValueMultiSeqs(unittest.TestCase):
inputs["min_tokens"][:] = 0 inputs["min_tokens"][:] = 0
outputs = self._run_and_get(inputs) outputs = self._run_and_get(inputs)
self._check_all_outputs(inputs, outputs) self._check_all_outputs(inputs, outputs)
# Match at accept_idx=2, loop increments to 3 # Match at accept_idx=1, loop increments to 2
self.assertEqual(outputs["accept_num"][0], 3) self.assertEqual(outputs["accept_num"][0], 2)
self.assertEqual(outputs["accept_tokens"][0, 2], -1) self.assertEqual(outputs["accept_tokens"][0, 1], inputs["end_ids"][0])
def test_match_in_pre_ids_only(self): def test_match_in_pre_ids_only(self):
"""Stop seq found entirely within token_ids_all (pre_ids), matching at accept_idx=0.""" """Stop seq found entirely within token_ids_all (pre_ids), matching at accept_idx=0."""
@@ -314,29 +316,29 @@ class TestSpeculateSetStopValueMultiSeqs(unittest.TestCase):
accept_tokens_len=5, accept_tokens_len=5,
max_model_len=32, max_model_len=32,
stop_seqs_bs=1, stop_seqs_bs=1,
stop_seqs_max_len=3, stop_seqs_max_len=4, # 需要4个元素
seed=30, seed=30,
) )
inputs["prompt_lens"][:] = 0 inputs["prompt_lens"][:] = 0
inputs["step_idx"][:] = 8 inputs["step_idx"][:] = 8
inputs["accept_num"][:] = 3 inputs["accept_num"][:] = 3
# pre_ids at step_idx positions: token_ids_all[0, 6]=50, [0,7]=60, [0,8]=70 # stop_seq partially in pre_ids, partially in accept_tokens
# stop_seq = [50, 60, 70], all 3 tokens are in pre_ids # For accept_idx=1: step_idx_now - accept_num + 1 + 1 = 8-3+1+1 = 7 >= stop_seq_len=4, so we check
# For accept_idx=0: step_idx_now + 0 + 1 = 9 >= stop_seq_len=3, so we check # i=3: stop_seq_len-1-i=0 <= accept_idx(1) -> accept_tokens[1-0] = accept_tokens[1] = 22
# i=2: pre_ids_idx = 8+0-(3-1-2) = 8 -> pre_ids_now[8] = 70 # i=2: stop_seq_len-1-i=1 <= accept_idx(1) -> accept_tokens[1-1] = accept_tokens[0] = 11
# i=1: pre_ids_idx = 8+0-(3-1-1) = 7 -> pre_ids_now[7] = 60 # i=1: stop_seq_len-1-i=2 > accept_idx(1) -> pre_ids_idx = 8-3+1-(4-1-1) = 6-2 = 4 -> pre_ids[4] = 60
# i=0: pre_ids_idx = 8+0-(3-1-0) = 6 -> pre_ids_now[6] = 50 # i=0: stop_seq_len-1-i=3 > accept_idx(1) -> pre_ids_idx = 8-3+1-(4-1-0) = 6-3 = 3 -> pre_ids[3] = 50
inputs["token_ids_all"][0, 6] = 50 inputs["token_ids_all"][0, 3] = 50
inputs["token_ids_all"][0, 7] = 60 inputs["token_ids_all"][0, 4] = 60
inputs["token_ids_all"][0, 8] = 70 inputs["accept_tokens"][0, :3] = [11, 22, 3]
inputs["accept_tokens"][0, :3] = [1, 2, 3] inputs["stop_seqs"][0, 0, :4] = [50, 60, 11, 22]
inputs["stop_seqs"][0, 0, :3] = [50, 60, 70] inputs["stop_seqs_len"][0, 0] = 4
inputs["stop_seqs_len"][0, 0] = 3
inputs["stop_flags"][:] = False inputs["stop_flags"][:] = False
inputs["min_tokens"][:] = 0 inputs["min_tokens"][:] = 0
outputs = self._run_and_get(inputs) outputs = self._run_and_get(inputs)
self._check_all_outputs(inputs, outputs) self._check_all_outputs(inputs, outputs)
self.assertEqual(outputs["accept_num"][0], 1) # Match at accept_idx=1, loop increments to 2
self.assertEqual(outputs["accept_num"][0], 2)
def test_already_stopped(self): def test_already_stopped(self):
"""Kernel skips sequences with stop_flags=True.""" """Kernel skips sequences with stop_flags=True."""