From a3ecfd23c3cf80420f8eb08d2c6e06a77020a41f Mon Sep 17 00:00:00 2001 From: Xin Li <119016172+xinli-sw@users.noreply.github.com> Date: Sun, 16 Aug 2026 22:48:10 -0400 Subject: [PATCH 1/5] Add acceptance length reference values for deepseek model This YAML file contains acceptance length reference values measured with SPEED-Bench for the deepseek-v4-pro-0813 model, detailing the golden AL values per number of speculative tokens. --- .../deepseek-v4-pro-0813-dspark.yaml | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) create mode 100644 golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml diff --git a/golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml b/golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml new file mode 100644 index 000000000..cab002f9d --- /dev/null +++ b/golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml @@ -0,0 +1,19 @@ +# Source GitHub Actions run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31742838308 +# Acceptance Length (AL) reference values measured with SPEED-Bench. +# dataset: coding | temperature: 1.0 | output_len: 4096 +# thinking_on chat_template_kwargs: {"thinking": true, "reasoning_effort": "high"} +# speculative-config: method=dspark | draft_sample_method=probabilistic +# Measured on deepseek-v4-pro-0813 (B300, vLLM DSpark), per num_speculative_tokens. +# Auto-generated by benchmarks/single_node/speedbench/dsv4dspark_fp4_b300_vllm.sh (speedbench-al.yml). +# +# key = num_speculative_tokens (DSpark level); value = golden AL +deepseek-v4-pro-0813: + thinking_on: + 1: 1.84 + 2: 2.51 + 3: 3.01 + 4: 3.36 + 5: 3.61 + 6: 3.77 + 7: 3.73 + 8: 3.47 From f1954358e44afc9cd06faea8706a4797f776cc3b Mon Sep 17 00:00:00 2001 From: Xin Li <119016172+xinli-sw@users.noreply.github.com> Date: Sun, 16 Aug 2026 22:49:06 -0400 Subject: [PATCH 2/5] Add dsv4-pro-0813-dspark.yaml file --- ...deepseek-v4-pro-0813-dspark.yaml => dsv4-pro-0813-dspark.yaml} | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename golden_al_distribution/{deepseek-v4-pro-0813-dspark.yaml => dsv4-pro-0813-dspark.yaml} (100%) diff --git a/golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml b/golden_al_distribution/dsv4-pro-0813-dspark.yaml similarity index 100% rename from golden_al_distribution/deepseek-v4-pro-0813-dspark.yaml rename to golden_al_distribution/dsv4-pro-0813-dspark.yaml From 8862c8f32ce57cb5f376df9c9c5c8513172ddfe9 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 17 Aug 2026 22:04:09 -0400 Subject: [PATCH 3/5] Promote dsv4 DSpark to agreed PoR and wire AL lookup key MODELS.md/MODELS_zh.md: move DeepSeek-V4-Pro-DSpark from proposed to agreed draft model (PoR) for agentic coding, now that golden AL reference values are committed for it. inject_synthetic_acceptance.py: add dsv4dspark/dsv4dsparkprob model-prefix mappings to the deepseek-v4-pro-0813 yaml key so future AgentX configs resolve the new DSpark curve instead of colliding with the dsv4 MTP key. Co-Authored-By: Claude Sonnet 5 --- MODELS.md | 2 +- MODELS_zh.md | 2 +- runners/inject_synthetic_acceptance.py | 4 ++++ 3 files changed, 6 insertions(+), 2 deletions(-) diff --git a/MODELS.md b/MODELS.md index c2ee6aa8b..2ddb69931 100644 --- a/MODELS.md +++ b/MODELS.md @@ -108,7 +108,7 @@ The table also records both the agreed plan-of-record (PoR) draft-model mapping | Model | Primary native/upstream engines | Agreed draft model(s) (PoR) | Proposed draft model(s) pending partner alignment | Additional engines | |---|---|---|---|---| -| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP | `deepseek-ai/DeepSeek-V4-Pro-DSpark`, proposed for AgentX only under the same synthetic-acceptance methodology and pending partner alignment. Single-turn 8k1k remains on the native MTP heads. | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | +| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP (Single-turn 8k1k); `deepseek-ai/DeepSeek-V4-Pro-DSpark` (Agentic coding only, under the same synthetic-acceptance methodology as Kimi-K3's DSpark PoR) | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | Kimi-K3 (`kimik3`) | native/upstream vLLM engine | `Inferact/Kimi-K3-DSpark` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | MiniMax-M3 (`minimaxm3`) | native/upstream vLLM engine | `Inferact/MiniMax-M3-EAGLE3` and/or `Inferact/MiniMax-M3-EAGLE3-GQA` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | GLM-5.2 (`glm5.2`) | native/upstream SGLang engine | native MTP | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | diff --git a/MODELS_zh.md b/MODELS_zh.md index d82628808..447a4a7b0 100644 --- a/MODELS_zh.md +++ b/MODELS_zh.md @@ -108,7 +108,7 @@ InferenceX 支持 SGLang 和 vLLM 双方的维护者，并响应 AI 实验室和 | 模型 | 首选原生/上游引擎 | 已达成一致的草稿模型（PoR） | 待合作伙伴对齐的草稿模型提案 | 其他引擎 | |---|---|---|---|---| -| DeepSeek-V4-Pro 1.6T（`dsv4`） | 原生/上游 vLLM 引擎和原生/上游 SGLang 引擎 | 原生 MTP | `deepseek-ai/DeepSeek-V4-Pro-DSpark`，仅提议用于 AgentX，并须遵循相同的合成接受方法，尚待合作伙伴对齐。单轮 8k1k 继续使用原生 MTP 头。 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | +| DeepSeek-V4-Pro 1.6T（`dsv4`） | 原生/上游 vLLM 引擎和原生/上游 SGLang 引擎 | 原生 MTP（单轮 8k1k）；`deepseek-ai/DeepSeek-V4-Pro-DSpark`（仅用于智能体编码，遵循与 Kimi-K3 DSpark PoR 相同的合成接受方法） | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | Kimi-K3（`kimik3`） | 原生/上游 vLLM 引擎 | `Inferact/Kimi-K3-DSpark` | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | MiniMax-M3（`minimaxm3`） | 原生/上游 vLLM 引擎 | `Inferact/MiniMax-M3-EAGLE3` 和/或 `Inferact/MiniMax-M3-EAGLE3-GQA` | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | GLM-5.2（`glm5.2`） | 原生/上游 SGLang 引擎 | 原生 MTP | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | diff --git a/runners/inject_synthetic_acceptance.py b/runners/inject_synthetic_acceptance.py index 80d2740e7..bafda7a04 100644 --- a/runners/inject_synthetic_acceptance.py +++ b/runners/inject_synthetic_acceptance.py @@ -40,6 +40,10 @@ MODEL_PREFIX_TO_YAML_KEY = { "dsv4": "deepseek-v4-pro", "dsr1": "deepseek-r1", + # DSpark ships as its own checkpoint (deepseek-ai/DeepSeek-V4-Pro-DSpark, + # dated 0813), distinct from the plain MTP checkpoint above. + "dsv4dspark": "deepseek-v4-pro-0813", + "dsv4dsparkprob": "deepseek-v4-pro-0813", } From 1d2b419d7ec998b8f22c832c999f804b438c3b98 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Tue, 18 Aug 2026 01:25:21 -0400 Subject: [PATCH 4/5] Fix dsv4 DSpark PoR checkpoint name to DeepSeek-V4-Pro-0813 Co-Authored-By: Claude Sonnet 5 --- MODELS.md | 2 +- MODELS_zh.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/MODELS.md b/MODELS.md index 2ddb69931..aa394a3f3 100644 --- a/MODELS.md +++ b/MODELS.md @@ -108,7 +108,7 @@ The table also records both the agreed plan-of-record (PoR) draft-model mapping | Model | Primary native/upstream engines | Agreed draft model(s) (PoR) | Proposed draft model(s) pending partner alignment | Additional engines | |---|---|---|---|---| -| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP (Single-turn 8k1k); `deepseek-ai/DeepSeek-V4-Pro-DSpark` (Agentic coding only, under the same synthetic-acceptance methodology as Kimi-K3's DSpark PoR) | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | +| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP (Single-turn 8k1k); `deepseek-ai/DeepSeek-V4-Pro-0813` (Agentic coding only, under the same synthetic-acceptance methodology as Kimi-K3's DSpark PoR) | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | Kimi-K3 (`kimik3`) | native/upstream vLLM engine | `Inferact/Kimi-K3-DSpark` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | MiniMax-M3 (`minimaxm3`) | native/upstream vLLM engine | `Inferact/MiniMax-M3-EAGLE3` and/or `Inferact/MiniMax-M3-EAGLE3-GQA` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | GLM-5.2 (`glm5.2`) | native/upstream SGLang engine | native MTP | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | diff --git a/MODELS_zh.md b/MODELS_zh.md index 447a4a7b0..f1b81927a 100644 --- a/MODELS_zh.md +++ b/MODELS_zh.md @@ -108,7 +108,7 @@ InferenceX 支持 SGLang 和 vLLM 双方的维护者，并响应 AI 实验室和 | 模型 | 首选原生/上游引擎 | 已达成一致的草稿模型（PoR） | 待合作伙伴对齐的草稿模型提案 | 其他引擎 | |---|---|---|---|---| -| DeepSeek-V4-Pro 1.6T（`dsv4`） | 原生/上游 vLLM 引擎和原生/上游 SGLang 引擎 | 原生 MTP（单轮 8k1k）；`deepseek-ai/DeepSeek-V4-Pro-DSpark`（仅用于智能体编码，遵循与 Kimi-K3 DSpark PoR 相同的合成接受方法） | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | +| DeepSeek-V4-Pro 1.6T（`dsv4`） | 原生/上游 vLLM 引擎和原生/上游 SGLang 引擎 | 原生 MTP（单轮 8k1k）；`deepseek-ai/DeepSeek-V4-Pro-0813`（仅用于智能体编码，遵循与 Kimi-K3 DSpark PoR 相同的合成接受方法） | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | Kimi-K3（`kimik3`） | 原生/上游 vLLM 引擎 | `Inferact/Kimi-K3-DSpark` | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | MiniMax-M3（`minimaxm3`） | 原生/上游 vLLM 引擎 | `Inferact/MiniMax-M3-EAGLE3` 和/或 `Inferact/MiniMax-M3-EAGLE3-GQA` | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | | GLM-5.2（`glm5.2`） | 原生/上游 SGLang 引擎 | 原生 MTP | 无 | 按照上述提交顺序指南及例外处理的其他非 vLLM/SGLang 引擎 | From 511825d224ded921e231a8ea47ddbee126731745 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Tue, 18 Aug 2026 14:45:42 -0400 Subject: [PATCH 5/5] Update DeepSeek-V4-Pro description in MODELS.md --- MODELS.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/MODELS.md b/MODELS.md index aa394a3f3..b62e8091f 100644 --- a/MODELS.md +++ b/MODELS.md @@ -108,7 +108,7 @@ The table also records both the agreed plan-of-record (PoR) draft-model mapping | Model | Primary native/upstream engines | Agreed draft model(s) (PoR) | Proposed draft model(s) pending partner alignment | Additional engines | |---|---|---|---|---| -| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP (Single-turn 8k1k); `deepseek-ai/DeepSeek-V4-Pro-0813` (Agentic coding only, under the same synthetic-acceptance methodology as Kimi-K3's DSpark PoR) | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | +| DeepSeek-V4-Pro 1.6T (`dsv4`) | native/upstream vLLM engine and native/upstream SGLang engine | native MTP (Single-turn 8k1k); native DSpark heads `deepseek-ai/DeepSeek-V4-Pro-0813` (Agentic coding only, under the same synthetic-acceptance methodology as Kimi-K3's DSpark PoR) | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | Kimi-K3 (`kimik3`) | native/upstream vLLM engine | `Inferact/Kimi-K3-DSpark` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | MiniMax-M3 (`minimaxm3`) | native/upstream vLLM engine | `Inferact/MiniMax-M3-EAGLE3` and/or `Inferact/MiniMax-M3-EAGLE3-GQA` | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above | | GLM-5.2 (`glm5.2`) | native/upstream SGLang engine | native MTP | None | Additional non-vLLM/SGLang engines under the ordering guideline and exceptions above |