Skip to content

[TRTLLM-11508][refactor] decouple MTP num_nextn_predict_layers from max_draft_len #12341

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
QiJune merged 3 commits into from
May 8, 2026
Merged

[TRTLLM-11508][refactor] decouple MTP num_nextn_predict_layers from max_draft_len #12341

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d586f50
[TRTLLM-11508][refactor] decouple MTP num_nextn_predict_layers from m…
zhaoyangwang-nvidia Apr 10, 2026
81620e5
[TRTLLM-11508][refactor] add backward compat for deprecated MTPDecodi…
zhaoyangwang-nvidia Apr 10, 2026
e136f55
Fix new ci issue
zhaoyangwang-nvidia Apr 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 1 addition & 2 deletions docs/source/features/speculative-decoding.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,8 +114,7 @@ MTP is currently only supported by Deepseek. MTP can be tuned with the following
```python
from tensorrt_llm.llmapi import MTPDecodingConfig

speculative_config = MTPDecodingConfig(
max_draft_len=3, num_nextn_predict_layers=3)
speculative_config = MTPDecodingConfig(max_draft_len=3)

llm = LLM("/path/to/deepseek_model", speculative_config=speculative_config)
```
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/curated/deepseek-r1-deepgemm.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ kv_cache_config:
stream_interval: 10
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
moe_config:
backend: DEEPGEMM
max_num_tokens: 3200
2 changes: 1 addition & 1 deletion examples/configs/curated/deepseek-r1-latency.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
use_relaxed_acceptance_for_thinking: true
relaxed_topk: 10
relaxed_delta: 0.6
2 changes: 1 addition & 1 deletion examples/configs/curated/deepseek-r1-throughput.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,4 +12,4 @@ kv_cache_config:
stream_interval: 10
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc1024.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc128.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc16.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc2048.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc256.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc32.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc512.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc64.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k1k_tp8_conc8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc1024.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc128.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc16.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc2048.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc256.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc32.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc512.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc64.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: DEEPGEMM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/1k8k_tp8_conc8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc1024.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc128.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ attention_dp_config:
timeout_iters: 100
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc16.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc2048.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 100
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc256.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 100
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc32.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 100
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc64.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 100
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/B200/8k1k_tp8_conc8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: TRTLLM
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/H200/1k1k_tp8_conc1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: CUTLASS
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/H200/1k1k_tp8_conc1024.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/H200/1k1k_tp8_conc128.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ attention_dp_config:
timeout_iters: 60
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 1
max_draft_len: 1
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/H200/1k1k_tp8_conc16.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: CUTLASS
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
2 changes: 1 addition & 1 deletion examples/configs/database/deepseek-ai/DeepSeek-R1-0528/H200/1k1k_tp8_conc2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ moe_config:
backend: CUTLASS
speculative_config:
decoding_type: MTP
num_nextn_predict_layers: 3
max_draft_len: 3
tensor_parallel_size: 8
moe_expert_parallel_size: 8
trust_remote_code: true
Expand Down
Loading
Loading