Skip to content

[Refactor][Quantization]KVCache quantization dtype specified by the --kv-quant-dtype parameter #15514

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
kunpengW-code merged 2 commits into from
Sep 19, 2026
Merged

[Refactor][Quantization]KVCache quantization dtype specified by the --kv-quant-dtype parameter #15514

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bb630a0
add --kv-cache-dtype argument
lcfenglinwan Sep 13, 2026
0c77390
fix ut
lcfenglinwan Sep 18, 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
10 changes: 10 additions & 0 deletions docs/source/tutorials/models/DeepSeek-V4-Flash.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,6 +176,7 @@ Single-node deployment completes both Prefill and Decode within the same node. T
--quantization ascend \
--port 8000 \
--block-size 128 \
--attention_config.indexer_kv_dtype int8 \
--speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '
Expand Down Expand Up @@ -219,6 +220,7 @@ Single-node deployment completes both Prefill and Decode within the same node. T
--quantization ascend \
--port 8000 \
--block-size 128 \
--attention_config.indexer_kv_dtype int8 \
--speculative-config '{"method": "dspark", "num_speculative_tokens": 7, "enforce_eager": true}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
```
Expand Down Expand Up @@ -257,6 +259,7 @@ Single-node deployment completes both Prefill and Decode within the same node. T
--quantization ascend \
--port 8000 \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '
Expand Down Expand Up @@ -300,6 +303,7 @@ Single-node deployment completes both Prefill and Decode within the same node. T
--quantization ascend \
--port 8000 \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"enforce_eager":true}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '{
Expand Down Expand Up @@ -533,6 +537,7 @@ Before you start, please:
--speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
--trust-remote-code \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--tokenizer-mode deepseek_v4 \
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
Expand Down Expand Up @@ -598,6 +603,7 @@ Before you start, please:
--max-num-batched-tokens 120 \
--max-num-seqs 60 \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--no-disable-hybrid-kv-cache-manager \
--no-enable-prefix-caching \
--trust-remote-code \
Expand Down Expand Up @@ -684,6 +690,7 @@ Before you start, please:
--speculative-config '{"num_speculative_tokens": 5,"method": "dspark","enforce_eager": true}' \
--trust-remote-code \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--tokenizer-mode deepseek_v4 \
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
Expand Down Expand Up @@ -750,6 +757,7 @@ Before you start, please:
--max-num-seqs 60 \
--async-scheduling \
--block-size 32 \
--attention_config.indexer_kv_dtype int8 \
--no-disable-hybrid-kv-cache-manager \
--no-enable-prefix-caching \
--trust-remote-code \
Expand Down Expand Up @@ -941,6 +949,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--additional-config '{"enable_cpu_binding": true, "enable_shared_expert_dp": true}' \
--speculative-config '{"num_speculative_tokens": 1, "method": "mtp","enforce_eager": true}' \
--kv-transfer-config \
Expand Down Expand Up @@ -1017,6 +1026,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--speculative-config '{"num_speculative_tokens": 1, "method": "mtp","enforce_eager": true}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--kv-transfer-config \
Expand Down
12 changes: 12 additions & 0 deletions docs/source/tutorials/models/DeepSeek-V4-Pro.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,6 +193,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config '{
"enable_multithread_load": true,
"num_threads": 128
Expand Down Expand Up @@ -270,6 +271,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config '{
"enable_multithread_load": true,
"num_threads": 128
Expand Down Expand Up @@ -341,6 +343,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--quantization ascend \
--speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
Expand Down Expand Up @@ -400,6 +403,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--quantization ascend \
--speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
Expand Down Expand Up @@ -466,6 +470,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--quantization ascend \
--speculative-config '{"num_speculative_tokens":5,"method":"dspark","enforce_eager":true}' \
Expand Down Expand Up @@ -529,6 +534,7 @@ The quantized model `DeepSeek-V4-Pro-w4a8-mtp` requires at least 2 Atlas 800 A3
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--quantization ascend \
--speculative-config '{"num_speculative_tokens":5,"method":"dspark","enforce_eager":true}' \
Expand Down Expand Up @@ -770,6 +776,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down Expand Up @@ -842,6 +849,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down Expand Up @@ -941,6 +949,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down Expand Up @@ -1006,6 +1015,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down Expand Up @@ -1214,6 +1224,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down Expand Up @@ -1291,6 +1302,7 @@ Before you start, please:
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--attention_config.indexer_kv_dtype int8 \
--model-loader-extra-config='{"enable_multithread_load": true, "num_threads": 128}' \
--trust-remote-code \
--quantization ascend \
Expand Down
Loading
Loading