From bfe0aa9219518b2a72a265fdaf4e1fe85ada9ae7 Mon Sep 17 00:00:00 2001
From: yiminghub2024 <482890@qq.com>
Date: Sat, 15 Aug 2026 21:37:59 +0800
Subject: [PATCH 1/3] Update models.md

---
 docs/recipes/models.md | 30 +++++++++++++++++++++++++++++-
 1 file changed, 29 insertions(+), 1 deletion(-)

diff --git a/docs/recipes/models.md b/docs/recipes/models.md
index 167de2f396..ef2ed61e8a 100644
--- a/docs/recipes/models.md
+++ b/docs/recipes/models.md
@@ -430,7 +430,7 @@ Notes:
 - The DeepEP layout pulls ahead from mid batch up, where its expert kernels and
   the second attention replica both pay off.
 
-### Qwen3.8-27B
+### Qwen3.8-27B(Blackwell)
 
 A dense 27B-class Qwen3.8 FP8 checkpoint on a single GPU, with self-speculative
 MTP (the draft model path points at the same checkpoint):
@@ -455,6 +455,34 @@ tokenspeed serve Qwen/Qwen3.8-27B-FP8 \
   --host 0.0.0.0 --port 8000
 ```
 
+### Qwen3.8-27B(Hopper)
+
+A dense 27B-class Qwen3.8 FP8 checkpoint on a single GPU, with self-speculative
+MTP (the draft model path points at the same checkpoint):
+
+```bash
+tokenspeed serve Qwen/Qwen3.8-27B-FP8 \
+  --served-model-name Qwen/Qwen3.8-27B-FP8 \
+  --world-size 1 \
+  --gpu-memory-utilization 0.9 \
+  --moe-backend fa3 \
+  --drafter-attention-backend fa3 \
+  --chunked-prefill-size 8192 \
+  --max-model-len 256000 \
+  --max-num-seqs 64 \
+  --kv-cache-dtype fp8_e4m3 \
+  --speculative-algorithm MTP \
+  --speculative-draft-model-path Qwen/Qwen3.8-27B-FP8 \
+  --speculative-num-steps 3 \
+  --speculative-eagle-topk 1 \
+  --speculative-num-draft-tokens 4 \
+  --reasoning-parser qwen3 \
+  --tool-call-parser qwen \
+  --disable-kvstore \
+  --host 0.0.0.0 --port 8000
+```
+
+
 ## GPT-OSS 20B / 120B
 
 Small GPT-OSS launches can start simple. Large GPT-OSS launches usually tune

From e115788b59c7df0466fadc3eb2e1924bc4492ce1 Mon Sep 17 00:00:00 2001
From: yiminghub2024 <482890@qq.com>
Date: Sat, 15 Aug 2026 21:47:03 +0800
Subject: [PATCH 2/3] Update getting-started.md with kernel installation
 support hopper

Update getting-started.md with kernel installation support hopper
it tested on h20
---
 docs/guides/getting-started.md | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/docs/guides/getting-started.md b/docs/guides/getting-started.md
index 7729ffb4aa..2d981843f0 100644
--- a/docs/guides/getting-started.md
+++ b/docs/guides/getting-started.md
@@ -45,12 +45,20 @@ pip install -e "./python" --no-build-isolation
 ```
 
 Install the kernel package. Its Python package metadata installs the selected
-backend dependencies automatically.
+backend dependencies automatically(blackwell).
 
 ```bash
 pip install -e tokenspeed-kernel/python/ --no-build-isolation
 ```
 
+Install the kernel package. Its Python package metadata installs the selected
+backend dependencies automatically(hopper).
+
+```bash
+TOKENSPEED_CUDA_ARCH=90a MAX_JOBS=32 \
+  pip install -e tokenspeed-kernel/python/ --no-build-isolation
+```
+
 Install the scheduler package:
 
 ```bash

From 776037c4336aabeaad45c3ece2f1cb3bf3c639ff Mon Sep 17 00:00:00 2001
From: yiminghub2024 <482890@qq.com>
Date: Sun, 16 Aug 2026 07:24:17 +0800
Subject: [PATCH 3/3] Remove moe-backend option from tokenspeed serve command

Removed the '--moe-backend fa3' option from the command.
---
 docs/recipes/models.md | 1 -
 1 file changed, 1 deletion(-)

diff --git a/docs/recipes/models.md b/docs/recipes/models.md
index ef2ed61e8a..ecaae18bb7 100644
--- a/docs/recipes/models.md
+++ b/docs/recipes/models.md
@@ -465,7 +465,6 @@ tokenspeed serve Qwen/Qwen3.8-27B-FP8 \
   --served-model-name Qwen/Qwen3.8-27B-FP8 \
   --world-size 1 \
   --gpu-memory-utilization 0.9 \
-  --moe-backend fa3 \
   --drafter-attention-backend fa3 \
   --chunked-prefill-size 8192 \
   --max-model-len 256000 \