spele1100

GLM-5.3-Flash-AWQ-INT4

Deploy Dedicated

README

License: apache-2.0

Model

Base: GLM-5.3-Flash (glm5_next arch, 45 layers, hc_mult=4, 288 experts, inter=2048, h=4096)
Quantization: AWQ INT4

Required vLLM build

Requires a custom vLLM fork (sm89 long-context fixes) — stock vLLM has a broken mHC fused-norm kernel on this arch (all-45-layer poisoning) and long-context CUDA-graph crashes.
Version matters — verify your build includes both the mHC kernel fix and the long-context graph fix before debugging anything else.
Env reference: Python 3.10, transformers 5.16.1, flashinfer + sm89-compatible indexer.

Bring-up (512K production)

bash

# 1. Verify your vLLM build includes the sm89 long-context fixes # 2. Launch under tmux (NEVER bare — a dying shell takes the server with it)tmux new -s vllm<your serve script># detach: Ctrl-b d ; reattach: tmux attach # 3. Health checks (loading ~46 safetensors shards takes several minutes)curl -s http://localhost:<port>/v1/models | head# smoke:curl -s http://localhost:<port>/v1/chat/completions -H 'Content-Type: application/json' \  -d '{"model":"glm53-flash-awq","messages":[{"role":"user","content":"用一句話介紹自己"}],"max_tokens":2048}'

Reference serve script:

bash

export VLLM_USE_BREAKABLE_CUDAGRAPH=0export PYTHONPATH=<path to your vLLM fork>:$PYTHONPATHexec vllm serve <model_path> \  --served-model-name glm53-flash-awq \  --tensor-parallel-size 8 \  --host 0.0.0.0 --port <port> \  --max-model-len 524288 \  --no-enable-prefix-caching \  --hf-overrides '{"text_config": {"index_topk": null, "index_n_heads": null, "index_head_dim": null, "index_kpool": null}}' \  --compilation-config "{\"cudagraph_mode\": \"FULL_DECODE_ONLY\"}" \  --gpu-memory-utilization 0.94 \  --max-num-batched-tokens 4096 \  --reasoning-parser glm47 \  --tool-call-parser glm47 \  --enable-auto-tool-choice

Key config notes:

Full-MLA mode: the hf-overrides null out the sparse indexer (index_topk etc.) — sparse retrieval is OFF in production. An sm89 sparse-indexer port can pass correctness in chat-mode probes, but very-long-context (1M) sparse is blocked by an upstream persistent_topk smem bug.
Graph mode: runs WITHOUT --enforce-eager — FULL_DECODE_ONLY cudagraph. (Enforce-eager was needed on older builds due to long-context CUDA-graph crashes; current fixes remove that requirement.)
Prefix caching disabled; KV bf16 (fp8 unsupported, see below); TP8; ~34 GiB peak per GPU during load.

Restart quirk (known gotcha)

If the server just died, don't relaunch immediately — wait ~1 min and clear /dev/shm residue (stale vLLM shm blocks cause the relaunch to fail). Then relaunch in the same tmux session.

Verification checklist

/v1/models responds; loading ~46 shards takes minutes, not seconds.
Chat smoke: greeting + Traditional Chinese + simple math — coherent, no repetitive/scaled-token signature. Use /v1/chat/completions or llm.chat() — never raw generate() (see lessons below). max_tokens ≥ 2048 (reasoning eats budget).
Needle test at ~400–500K context: verify retrieval inside the 512K window.
Thinking field: expect reasoning (or reasoning_content if fronted by a proxy that normalizes it). Should appear once the model actually reasons.

Context-length verdicts (verified 2026-08-30)

512K (524288): production-ready — this is the verified config.
≤ ~600K: usable.
1M: boots but retrieval is unreliable — do not trust answers at 1M.
Nothing above 512K is supported. 512K is the hard, tested ceiling of this release; longer contexts are known-broken territory.

Known limits & lessons (2026-08-29/30)

Old vLLM builds = guaranteed garbage. The mHC kernel fix must be present in your build — always verify before debugging.
Chat-model probes MUST use the chat template (llm.chat() / /v1/chat/completions). Raw generate() degenerates on this model and looks like a model bug but isn't.
fp8 KV cache NOT supported on sm89 MLA — vLLM init fails ("FP8 kv_data_type for MLA is only supported with the fa3 backend on SM90"). FA3-only. Keep KV bf16 (auto); do not retry.
pe_dim=64: GLM has rope=0 — fp8_ds_mla packed layout (SM120 path) can never work; zero-pad rope trick is required (already in the fix branches).
Chat template exposes reasoning inline in content on direct probes; use max_tokens ≥ 2048.
vLLM serve hot path runs inside CUDA graph capture — python hooks won't fire; use offline eager mode for probing.
Always launch under tmux; see restart quirk above.

Validation history

2026-08-29 smoke: greeting / math / Traditional Chinese — coherent; HF reference parity layers 0–2 cos 0.9999+; effective GEMM cos 0.992, ratio 1.008 (post fold-fix).
2026-08-30: 512K production config verified live; context verdicts above.

vLLM fork & build procedure (required — stock vLLM will NOT work)

The custom fork is mirrored at https://github.com/spele1100/vllm-glm53, branch sm89-longctx-fix-v2, verified commit f2935ce322.

⚠️ This is an unofficial, experimental fork — it is not part of official vLLM and receives no support or maintenance guarantee. It exists because stock vLLM does not properly support this architecture on sm89. Use at your own risk and always validate your build against the checklist below.

Patch history (5 commits on top of vLLM 0.28.1rc1.dev):

cdc814dc4 — allow MLADimensions(256,0,256) in FA prefill; skip indexer weights when index_topk=null
dc5fcfd3b — enable GLM-5.3 sparse indexer + sparse MLA decode on Ada (portable MQA fallbacks, FA2 nope MLA with zero-padded rope, JIT toolchain fixes)
1358bbd8c — mHC fused-norm correctness fix (the all-45-layer poisoning bug)
005c57e0d — bound indexer MQA logits memory; fix long-context prefill crashes
f2935ce32 — skip indexer checkpoint tensors when index_topk=null (no-indexer mode used in production)

Build steps

bash

git clone https://github.com/spele1100/vllm-glm53.git ~/llm/vllm-glm53cd ~/llm/vllm-glm53git checkout sm89-longctx-fix-v2git rev-parse --short=10 HEAD   # must be f2935ce322 or newer python3.10 -m venv ~/llm/venv-glm53-servesource ~/llm/venv-glm53-serve/bin/activatepip install -U pip wheel # transformers pinpip install 'transformers==5.16.1' # FlashInfer: pip wheel 0.6.18 is broken on sm89/sm121 — build from sourcegit clone https://github.com/flashinfer-ai/flashinfer.git /tmp/flashinferpip install -e /tmp/flashinfer --no-build-isolationexport FLASHINFER_DISABLE_VERSION_CHECK=1 # vLLM fork (editable, so PYTHONPATH pick-up works)pip install -e ~/llm/vllm-glm53 --no-build-isolation

Runtime env (must match the serve script)

bash

N=$HOME/llm/venv-glm53-serve/lib/python3.10/site-packages/nvidia/cu13export CUDA_HOME=$Nexport PATH=$N/bin:$PATHexport PYTHONPATH=$HOME/llm/vllm-glm53:$PYTHONPATHexport FLASHINFER_DISABLE_VERSION_CHECK=1export PYTHONFAULTHANDLER=1export VLLM_USE_BREAKABLE_CUDAGRAPH=0

Build/run gotchas (learned the hard way)

If vLLM init crashes in FlashInfer JIT, the version check masking is the usual suspect — keep FLASHINFER_DISABLE_VERSION_CHECK=1.
VLLM_USE_BREAKABLE_CUDAGRAPH=0 + FULL_DECODE_ONLY graph mode is the verified combo; do not mix with enforce-eager configs from older docs.
After any branch/commit change, re-run the verification checklist above — config- and checkpoint-version mismatches are the #1 cause of gibberish output reports.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

spele1100

Model Tree

Base

zai-org/GLM-5.3-Flash

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model

Base: GLM-5.3-Flash (glm5_next arch, 45 layers, hc_mult=4, 288 experts, inter=2048, h=4096)
Quantization: AWQ INT4

Required vLLM build

Requires a custom vLLM fork (sm89 long-context fixes) — stock vLLM has a broken mHC fused-norm kernel on this arch (all-45-layer poisoning) and long-context CUDA-graph crashes.
Version matters — verify your build includes both the mHC kernel fix and the long-context graph fix before debugging anything else.
Env reference: Python 3.10, transformers 5.16.1, flashinfer + sm89-compatible indexer.

Bring-up (512K production)

bash

# 1. Verify your vLLM build includes the sm89 long-context fixes # 2. Launch under tmux (NEVER bare — a dying shell takes the server with it)tmux new -s vllm<your serve script># detach: Ctrl-b d ; reattach: tmux attach # 3. Health checks (loading ~46 safetensors shards takes several minutes)curl -s http://localhost:<port>/v1/models | head# smoke:curl -s http://localhost:<port>/v1/chat/completions -H 'Content-Type: application/json' \  -d '{"model":"glm53-flash-awq","messages":[{"role":"user","content":"用一句話介紹自己"}],"max_tokens":2048}'

# 1. Verify your vLLM build includes the sm89 long-context fixes # 2. Launch under tmux (NEVER bare — a dying shell takes the server with it)tmux new -s vllm<your serve script># detach: Ctrl-b d ; reattach: tmux attach # 3. Health checks (loading ~46 safetensors shards takes several minutes)curl -s http://localhost:<port>/v1/models | head# smoke:curl -s http://localhost:<port>/v1/chat/completions -H 'Content-Type: application/json' \  -d '{"model":"glm53-flash-awq","messages":[{"role":"user","content":"用一句話介紹自己"}],"max_tokens":2048}'

Reference serve script:

bash

export VLLM_USE_BREAKABLE_CUDAGRAPH=0export PYTHONPATH=<path to your vLLM fork>:$PYTHONPATHexec vllm serve <model_path> \  --served-model-name glm53-flash-awq \  --tensor-parallel-size 8 \  --host 0.0.0.0 --port <port> \  --max-model-len 524288 \  --no-enable-prefix-caching \  --hf-overrides '{"text_config": {"index_topk": null, "index_n_heads": null, "index_head_dim": null, "index_kpool": null}}' \  --compilation-config "{\"cudagraph_mode\": \"FULL_DECODE_ONLY\"}" \  --gpu-memory-utilization 0.94 \  --max-num-batched-tokens 4096 \  --reasoning-parser glm47 \  --tool-call-parser glm47 \  --enable-auto-tool-choice

export VLLM_USE_BREAKABLE_CUDAGRAPH=0export PYTHONPATH=<path to your vLLM fork>:$PYTHONPATHexec vllm serve <model_path> \  --served-model-name glm53-flash-awq \  --tensor-parallel-size 8 \  --host 0.0.0.0 --port <port> \  --max-model-len 524288 \  --no-enable-prefix-caching \  --hf-overrides '{"text_config": {"index_topk": null, "index_n_heads": null, "index_head_dim": null, "index_kpool": null}}' \  --compilation-config "{\"cudagraph_mode\": \"FULL_DECODE_ONLY\"}" \  --gpu-memory-utilization 0.94 \  --max-num-batched-tokens 4096 \  --reasoning-parser glm47 \  --tool-call-parser glm47 \  --enable-auto-tool-choice

Key config notes:

Full-MLA mode: the hf-overrides null out the sparse indexer (index_topk etc.) — sparse retrieval is OFF in production. An sm89 sparse-indexer port can pass correctness in chat-mode probes, but very-long-context (1M) sparse is blocked by an upstream persistent_topk smem bug.
Graph mode: runs WITHOUT --enforce-eager — FULL_DECODE_ONLY cudagraph. (Enforce-eager was needed on older builds due to long-context CUDA-graph crashes; current fixes remove that requirement.)
Prefix caching disabled; KV bf16 (fp8 unsupported, see below); TP8; ~34 GiB peak per GPU during load.

Restart quirk (known gotcha)

If the server just died, don't relaunch immediately — wait ~1 min and clear /dev/shm residue (stale vLLM shm blocks cause the relaunch to fail). Then relaunch in the same tmux session.

Verification checklist

/v1/models responds; loading ~46 shards takes minutes, not seconds.
Chat smoke: greeting + Traditional Chinese + simple math — coherent, no repetitive/scaled-token signature. Use /v1/chat/completions or llm.chat() — never raw generate() (see lessons below). max_tokens ≥ 2048 (reasoning eats budget).
Needle test at ~400–500K context: verify retrieval inside the 512K window.
Thinking field: expect reasoning (or reasoning_content if fronted by a proxy that normalizes it). Should appear once the model actually reasons.

Context-length verdicts (verified 2026-08-30)

512K (524288): production-ready — this is the verified config.
≤ ~600K: usable.
1M: boots but retrieval is unreliable — do not trust answers at 1M.
Nothing above 512K is supported. 512K is the hard, tested ceiling of this release; longer contexts are known-broken territory.

Known limits & lessons (2026-08-29/30)

Old vLLM builds = guaranteed garbage. The mHC kernel fix must be present in your build — always verify before debugging.
Chat-model probes MUST use the chat template (llm.chat() / /v1/chat/completions). Raw generate() degenerates on this model and looks like a model bug but isn't.
fp8 KV cache NOT supported on sm89 MLA — vLLM init fails ("FP8 kv_data_type for MLA is only supported with the fa3 backend on SM90"). FA3-only. Keep KV bf16 (auto); do not retry.
pe_dim=64: GLM has rope=0 — fp8_ds_mla packed layout (SM120 path) can never work; zero-pad rope trick is required (already in the fix branches).
Chat template exposes reasoning inline in content on direct probes; use max_tokens ≥ 2048.
vLLM serve hot path runs inside CUDA graph capture — python hooks won't fire; use offline eager mode for probing.
Always launch under tmux; see restart quirk above.

Validation history

2026-08-29 smoke: greeting / math / Traditional Chinese — coherent; HF reference parity layers 0–2 cos 0.9999+; effective GEMM cos 0.992, ratio 1.008 (post fold-fix).
2026-08-30: 512K production config verified live; context verdicts above.

vLLM fork & build procedure (required — stock vLLM will NOT work)

The custom fork is mirrored at https://github.com/spele1100/vllm-glm53, branch sm89-longctx-fix-v2, verified commit f2935ce322.

⚠️ This is an unofficial, experimental fork — it is not part of official vLLM and receives no support or maintenance guarantee. It exists because stock vLLM does not properly support this architecture on sm89. Use at your own risk and always validate your build against the checklist below.

Patch history (5 commits on top of vLLM 0.28.1rc1.dev):

cdc814dc4 — allow MLADimensions(256,0,256) in FA prefill; skip indexer weights when index_topk=null
dc5fcfd3b — enable GLM-5.3 sparse indexer + sparse MLA decode on Ada (portable MQA fallbacks, FA2 nope MLA with zero-padded rope, JIT toolchain fixes)
1358bbd8c — mHC fused-norm correctness fix (the all-45-layer poisoning bug)
005c57e0d — bound indexer MQA logits memory; fix long-context prefill crashes
f2935ce32 — skip indexer checkpoint tensors when index_topk=null (no-indexer mode used in production)

Build steps

bash

git clone https://github.com/spele1100/vllm-glm53.git ~/llm/vllm-glm53cd ~/llm/vllm-glm53git checkout sm89-longctx-fix-v2git rev-parse --short=10 HEAD   # must be f2935ce322 or newer python3.10 -m venv ~/llm/venv-glm53-servesource ~/llm/venv-glm53-serve/bin/activatepip install -U pip wheel # transformers pinpip install 'transformers==5.16.1' # FlashInfer: pip wheel 0.6.18 is broken on sm89/sm121 — build from sourcegit clone https://github.com/flashinfer-ai/flashinfer.git /tmp/flashinferpip install -e /tmp/flashinfer --no-build-isolationexport FLASHINFER_DISABLE_VERSION_CHECK=1 # vLLM fork (editable, so PYTHONPATH pick-up works)pip install -e ~/llm/vllm-glm53 --no-build-isolation

git clone https://github.com/spele1100/vllm-glm53.git ~/llm/vllm-glm53cd ~/llm/vllm-glm53git checkout sm89-longctx-fix-v2git rev-parse --short=10 HEAD   # must be f2935ce322 or newer python3.10 -m venv ~/llm/venv-glm53-servesource ~/llm/venv-glm53-serve/bin/activatepip install -U pip wheel # transformers pinpip install 'transformers==5.16.1' # FlashInfer: pip wheel 0.6.18 is broken on sm89/sm121 — build from sourcegit clone https://github.com/flashinfer-ai/flashinfer.git /tmp/flashinferpip install -e /tmp/flashinfer --no-build-isolationexport FLASHINFER_DISABLE_VERSION_CHECK=1 # vLLM fork (editable, so PYTHONPATH pick-up works)pip install -e ~/llm/vllm-glm53 --no-build-isolation

Runtime env (must match the serve script)

bash

N=$HOME/llm/venv-glm53-serve/lib/python3.10/site-packages/nvidia/cu13export CUDA_HOME=$Nexport PATH=$N/bin:$PATHexport PYTHONPATH=$HOME/llm/vllm-glm53:$PYTHONPATHexport FLASHINFER_DISABLE_VERSION_CHECK=1export PYTHONFAULTHANDLER=1export VLLM_USE_BREAKABLE_CUDAGRAPH=0

N=$HOME/llm/venv-glm53-serve/lib/python3.10/site-packages/nvidia/cu13export CUDA_HOME=$Nexport PATH=$N/bin:$PATHexport PYTHONPATH=$HOME/llm/vllm-glm53:$PYTHONPATHexport FLASHINFER_DISABLE_VERSION_CHECK=1export PYTHONFAULTHANDLER=1export VLLM_USE_BREAKABLE_CUDAGRAPH=0

Build/run gotchas (learned the hard way)

If vLLM init crashes in FlashInfer JIT, the version check masking is the usual suspect — keep FLASHINFER_DISABLE_VERSION_CHECK=1.
VLLM_USE_BREAKABLE_CUDAGRAPH=0 + FULL_DECODE_ONLY graph mode is the verified combo; do not mix with enforce-eager configs from older docs.
After any branch/commit change, re-run the verification checklist above — config- and checkpoint-version mismatches are the #1 cause of gibberish output reports.