From 443f72087c6c6fa660b903422c254805de0a8a99 Mon Sep 17 00:00:00 2001 From: tutel-agent <211217074+tutel-agent@users.noreply.github.com> Date: Tue, 7 Jul 2026 21:36:42 +0800 Subject: [PATCH 1/5] Bug fix for 1M context on A100 80GB SXM --- README.md | 26 +++++++++++++++++++------- 1 file changed, 19 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index 932fb4967..c854a4952 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,4 @@ -# Tutel +image# Tutel Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parallel solution proposing ["No-penalty Parallism/Sparsity/Capacity/.. Switching"](https://mlsys.org/media/mlsys-2023/Slides/2477.pdf) for modern training and inference that have dynamic behaviors. @@ -10,7 +10,17 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > [!TIP] > #### Steps for GLM-5/5.1/5.2 (Claude-Code Mode): > +> ☑ A100/H100 80Gx8 (SXM): max-context-size = 1M +> +> ☑ MI300 192GBx8 (PCIe5): max-context-size = 1M × N +> > ```sh +> +---------------+--------------------+--------------------+ +> | | AIME-26 (No Tools) | AIME-26 (Tool Use) | +> +---------------+--------------------+--------------------+ +> | GLM-5.2 NVFP4 | 93.0% | 99.6% | +> +---------------+--------------------+--------------------+ +> > [Model Downloads] > pip3 install -U "huggingface_hub[cli]" --upgrade > hf download --local-dir nvidia/GLM-5.2-NVFP4 nvidia/GLM-5.2-NVFP4 @@ -21,7 +31,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > docker run -e WORKER=1 -e LOCAL_SIZE=8 -p 8000:8000 -it --rm --ipc=host --shm-size=8g \ > --ulimit memlock=-1 --ulimit stack=67108864 -v /:/host -w /host$(pwd) \ > -v /usr/lib/x86_64-linux-gnu/libcuda.so.1:/usr/lib/x86_64-linux-gnu/libcuda.so.1 --privileged \ -> tutelgroup/deepseek-671b:a100x8-chat-20260618 --serve=core \ +> tutelgroup/deepseek-671b:a100x8-chat-20260707 --serve=core \ > --try_path nvidia/GLM-5.2-NVFP4 \ > --try_path nvidia/GLM-5.1-NVFP4 \ > --try_path nvidia/GLM-5-NVFP4 \ @@ -31,7 +41,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > docker run -e WORKER=1 -e LOCAL_SIZE=8 -p 8000:8000 -it --rm --ipc=host --shm-size=8g \ > --ulimit memlock=-1 --ulimit stack=67108864 -v /:/host -w /host$(pwd) \ > --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --device=/dev/kfd --device=/dev/dri --group-add=video \ -> tutelgroup/deepseek-671b:mi300x8-chat-20260618 --serve=core \ +> tutelgroup/deepseek-671b:mi300x8-chat-20260707 --serve=core \ > --try_path nvidia/GLM-5.2-NVFP4 \ > --try_path nvidia/GLM-5.1-NVFP4 \ > --try_path nvidia/GLM-5-NVFP4 \ @@ -41,7 +51,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > #### Setup Claude Code for Linux / WSL (Ubuntu >= 24.04): > ```sh > sudo apt-get install -y npm -> sudo npm install -g @anthropic-ai/claude-code@2.1.90 +> sudo npm install -g @anthropic-ai/claude-code@2.1.197 > cat > run_claude.sh < mkdir -p config/ > export ANTHROPIC_BASE_URL="http://0.0.0.0:8000" @@ -91,7 +101,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > docker run -e LOCAL_SIZE=8 -e WORKER=1 -it --rm --ipc=host --net=host --shm-size=8g \ > --ulimit memlock=-1 --ulimit stack=67108864 -v /:/host -w /host$(pwd) -v /tmp:/tmp \ > -v /usr/lib/x86_64-linux-gnu/libcuda.so.1:/usr/lib/x86_64-linux-gnu/libcuda.so.1 --privileged \ -> tutelgroup/deepseek-671b:a100x8-chat-20260618 --serve=webui --listen_port 8000 \ +> tutelgroup/deepseek-671b:a100x8-chat-20260707 --serve=webui --listen_port 8000 \ > --try_path moonshotai/Kimi-K2.7-Code \ > --try_path moonshotai/Kimi-K2.6 \ > --try_path nvidia/Kimi-K2.5-NVFP4 \ @@ -104,7 +114,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > docker run -e LOCAL_SIZE=8 -e WORKER=1 -it --rm --ipc=host --net=host --shm-size=8g \ > --ulimit memlock=-1 --ulimit stack=67108864 --device=/dev/kfd --device=/dev/dri --group-add=video \ > --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /:/host -w /host$(pwd) -v /tmp:/tmp \ -> tutelgroup/deepseek-671b:mi300x8-chat-20260618 --serve=webui --listen_port 8000 \ +> tutelgroup/deepseek-671b:mi300x8-chat-20260707 --serve=webui --listen_port 8000 \ > --try_path moonshotai/Kimi-K2.7-Code \ > --try_path moonshotai/Kimi-K2.6 \ > --try_path nvidia/Kimi-K2.5-NVFP4 \ @@ -131,7 +141,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > hf download microsoft/VibeVoice-1.5B --local-dir microsoft/VibeVoice-1.5B > hf download Qwen/Qwen2.5-1.5B --local-dir Qwen/Qwen2.5-1.5B > -> hf download microsoft/VibeVoice-Large --local-dir aoi-ot/VibeVoice-Large +> hf download aoi-ot/VibeVoice-Large --local-dir microsoft/VibeVoice-Large > hf download Qwen/Qwen2.5-7B --local-dir Qwen/Qwen2.5-7B > > [Microsoft VibeVoice (ND_A100/H100/B200 only)] @@ -174,6 +184,8 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall ## What's New: +> Image-*20260707*: Memory fixes for 1M context on A100 80GB SXM. +> > Image-*20260618*: Fit GLM-5.2 1M context into A100 80GB x 8 SXM. > > Image-*20260603*: Improved Claude Code Tooling Performance for GLM-5/5.1. From 79a5c96b7bc929da2b18c1f26d0469458ebff41c Mon Sep 17 00:00:00 2001 From: tutel-agent <211217074+tutel-agent@users.noreply.github.com> Date: Tue, 7 Jul 2026 21:38:13 +0800 Subject: [PATCH 2/5] Update README.md --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c854a4952..137159d16 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,4 @@ -image# Tutel +# Tutel Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parallel solution proposing ["No-penalty Parallism/Sparsity/Capacity/.. Switching"](https://mlsys.org/media/mlsys-2023/Slides/2477.pdf) for modern training and inference that have dynamic behaviors. From d1037a79ebeec0270bdb6c62698b07228536332f Mon Sep 17 00:00:00 2001 From: tutel-agent <211217074+tutel-agent@users.noreply.github.com> Date: Tue, 7 Jul 2026 21:54:20 +0800 Subject: [PATCH 3/5] Update claude-code to 2.1.197 --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 137159d16..2de089f5a 100644 --- a/README.md +++ b/README.md @@ -69,7 +69,7 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > ```sh > winget install OpenJS.NodeJS.LTS > winget install --id Git.Git -e --source winget -> npm install -g @anthropic-ai/claude-code@2.1.90 +> npm install -g @anthropic-ai/claude-code@2.1.197 > ( > echo(@echo off > echo(if not exist config mkdir config From 1eb02520675a006a4ab036a450b8ada65963df1e Mon Sep 17 00:00:00 2001 From: tutel-agent <211217074+tutel-agent@users.noreply.github.com> Date: Wed, 8 Jul 2026 00:37:22 +0800 Subject: [PATCH 4/5] Update README.md --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 2de089f5a..04e104e88 100644 --- a/README.md +++ b/README.md @@ -10,9 +10,9 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > [!TIP] > #### Steps for GLM-5/5.1/5.2 (Claude-Code Mode): > -> ☑ A100/H100 80Gx8 (SXM): max-context-size = 1M +> ☑ A100x8/H100x8 (80G SXM): max-context-size = 1M > -> ☑ MI300 192GBx8 (PCIe5): max-context-size = 1M × N +> ☑ MI300x8 (192GB PCIe5): max-context-size = 1M × N > > ```sh > +---------------+--------------------+--------------------+ From b6455fcf679a14cc04b3b5caaea5d3c79b653d87 Mon Sep 17 00:00:00 2001 From: tutel-agent <211217074+tutel-agent@users.noreply.github.com> Date: Wed, 8 Jul 2026 13:36:59 +0800 Subject: [PATCH 5/5] Update README.md --- README.md | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 04e104e88..98bc9d2b1 100644 --- a/README.md +++ b/README.md @@ -16,9 +16,11 @@ Tutel MoE: An Optimized Mixture-of-Experts Implementation, also the first parall > > ```sh > +---------------+--------------------+--------------------+ -> | | AIME-26 (No Tools) | AIME-26 (Tool Use) | +> | [GLM-5.2] | (NVFP4 + No Tools) | (NVFP4 + Tool Use) | > +---------------+--------------------+--------------------+ -> | GLM-5.2 NVFP4 | 93.0% | 99.6% | +> | AIME-2025 | 96.1% | 99.5% | +> +---------------+--------------------+--------------------+ +> | AIME-2026 | 93.5% | 99.6% | > +---------------+--------------------+--------------------+ > > [Model Downloads]