Skip to content

fix: update VLLM_USE_DEEP_GEMM hub to default to 0 #293

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
velaraptor-runpod merged 2 commits into from
May 27, 2026
Merged

fix: update VLLM_USE_DEEP_GEMM hub to default to 0 #293

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
4c91f2c
fix: update VLLM_USE_DEEP_GEMM hub to default to 0
velaraptor-runpod May 27, 2026
9edc571
fix: update configuration.md
velaraptor-runpod May 27, 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
6 changes: 3 additions & 3 deletions .runpod/hub.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -810,9 +810,9 @@
"key": "VLLM_USE_DEEP_GEMM",
"input": {
"name": "Use DeepGEMM",
"type": "boolean",
"description": "Enable DeepGEMM FP8 kernels (MoE and MQA logits). Disabled by default — enable on H100/H200 for potential throughput gains. Some GPUs (e.g. H20) may perform better with this off. Enabling also adds a warmup period on startup.",
"default": false,
"type": "string",
"description": "Enable DeepGEMM FP8 kernels (MoE and MQA logits). Set to 1 to enable, 0 to disable. Required for DeepSeek V4 models. Disabled by default — enable on H100/H200 for potential throughput gains. Some GPUs (e.g. H20) may perform better with this off.",
"default": "0",
"advanced": true
}
}
Expand Down
4 changes: 2 additions & 2 deletions docs/configuration.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,12 +97,12 @@ If `SPECULATIVE_CONFIG` is set, it takes priority over individual env vars. When
| `MAX_SEQ_LEN_TO_CAPTURE` | `8192` | `int` | Maximum context length covered by CUDA graphs. When a sequence has context length larger than this, we fall back to eager mode. |
| `DISABLE_CUSTOM_ALL_REDUCE` | `0` | `int` | Enables or disables custom all reduce. |
| `ENABLE_EXPERT_PARALLEL` | `False` | `bool` | Enable Expert Parallel for MoE models. |
| `VLLM_USE_DEEP_GEMM` | `0` | `bool` | Enable DeepGEMM FP8 kernels for MoE and MQA logits computation. Disabled by default. See note below. |
| `VLLM_USE_DEEP_GEMM` | `0` | `str` (`0`/`1`) | Enable DeepGEMM FP8 kernels for MoE and MQA logits computation. Disabled by default. Must be `"0"` or `"1"` — not `true`/`false`. See note below. |
| `ATTENTION_BACKEND` | `None` | `str` | Attention backend to use (e.g., `FLASH_ATTN`, `FLASHINFER`, `TRITON_FLASH_ATTN`). Replaces deprecated `VLLM_ATTENTION_BACKEND`. |
| `ASYNC_SCHEDULING` | `None` | `bool` | Enable async scheduling (overlaps engine scheduling with GPU execution). Default: enabled in vLLM 0.14.0+. Set to `false` to disable. |
| `STREAM_INTERVAL` | `1` | `int` | Controls how often to yield streaming results. Lower = more frequent updates. |

> **Note (`VLLM_USE_DEEP_GEMM`):** DeepGEMM is used in two places: MoE weight computation and MQA logits computation. It is necessary for MQA logits computation on supported hardware. Set `VLLM_USE_DEEP_GEMM=0` to disable the MoE part and fall back to flashinfer/cutlass FP8 kernels. Some users report better performance with `VLLM_USE_DEEP_GEMM=0`, particularly on H20 GPUs. Disabling it also skips the DeepGEMM warmup phase, reducing cold-start time. Requires CUDA 13.0+ and SM90+ (H100/H200) to use; the library is installed but inactive by default.
> **Note (`VLLM_USE_DEEP_GEMM`):** DeepGEMM is used in two places: MoE weight computation and MQA logits computation. It is necessary for MQA logits computation on supported hardware — required for DeepSeek V4 models. Set `VLLM_USE_DEEP_GEMM=1` to enable. Set `VLLM_USE_DEEP_GEMM=0` to disable the MoE part and fall back to flashinfer/cutlass FP8 kernels. **Value must be `"0"` or `"1"` — not `"true"`/`"false"`.** Some users report better performance with `VLLM_USE_DEEP_GEMM=0`, particularly on H20 GPUs. Disabling it also skips the DeepGEMM warmup phase, reducing cold-start time. Requires CUDA 13.0+ and SM90+ (H100/H200) to use; the library is installed but inactive by default.

## Tokenizer Settings

Expand Down
Loading