Merge pull request #293 from runpod-workers/fix/update-deep-gemm-hub-value

fix: update VLLM_USE_DEEP_GEMM hub to default to 0

.runpod/README.md

@@ -6,7 +6,7 @@ Run LLMs using [vLLM](https://docs.vllm.ai) with an OpenAI-compatible API
 
 [![RunPod](https://api.runpod.io/badge/runpod-workers/worker-vllm)](https://www.runpod.io/console/hub/runpod-workers/worker-vllm)
 
-Current vLLM version: [0.19.1](https://github.com/vllm-project/vllm/releases/tag/v0.19.1)
+Current vLLM version: [0.20.2](https://github.com/vllm-project/vllm/releases/tag/v0.20.2)
 
 ---
 

.runpod/hub.json

@@ -9,7 +9,7 @@
     "containerDiskInGb": 150,
     "gpuIds": "ADA_80_PRO,AMPERE_80",
     "gpuCount": 1,
-    "allowedCudaVersions": ["12.9", "12.8"],
+    "allowedCudaVersions": ["13.0"],
     "presets": [
       {
         "name": "deepseek-ai/deepseek-r1-distill-llama-8b",
@@ -805,6 +805,16 @@
           "default": "expandable_segments:True",
           "advanced": true
         }
+      },
+      {
+        "key": "VLLM_USE_DEEP_GEMM",
+        "input": {
+          "name": "Use DeepGEMM",
+          "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
+        }
       }
     ]
   }

Dockerfile

@@ -1,16 +1,17 @@
-FROM nvidia/cuda:12.9.1-base-ubuntu22.04 
+FROM nvidia/cuda:13.0.2-devel-ubuntu22.04
 
 RUN apt-get update -y \
-    && apt-get install -y python3-pip curl \
+    && apt-get install -y python3-pip curl git \
     && curl -LsSf https://astral.sh/uv/install.sh  | sh
 
 ENV PATH="/root/.local/bin:$PATH"
 
-RUN ldconfig /usr/local/cuda-12.9/compat/
+RUN ldconfig /usr/local/cuda-13.0/compat/
 
-# Install vLLM with FlashInfer - use CUDA 12.9 PyTorch wheels
+# Install vLLM with FlashInfer - use CUDA 130 PyTorch wheels
 RUN uv pip install --system "packaging>=24.2" && \
-    uv pip install --system "vllm[flashinfer]==0.19.1" --extra-index-url https://download.pytorch.org/whl/cu129
+    uv pip install --system "vllm[flashinfer]==0.20.2" && \
+    uv pip install --system git+https://github.com/deepseek-ai/DeepGEMM.git@714dd1a4a980f7937a74343d19a8eba4fe321480 --no-build-isolation
 
 # Install additional Python dependencies (after vLLM to avoid PyTorch version conflicts)
 COPY builder/requirements.txt /requirements.txt
@@ -42,7 +43,9 @@ ENV MODEL_NAME=$MODEL_NAME \
     # Prevent rayon thread pool panic in containers where ulimit -u < nproc
     # (tokenizers uses Rust's rayon which tries to spawn threads = CPU cores)
     TOKENIZERS_PARALLELISM=false \
-    RAYON_NUM_THREADS=4
+    RAYON_NUM_THREADS=4 \
+    # Disable DeepGEMM MoE kernels by default; override with VLLM_USE_DEEP_GEMM=1 to enable
+    VLLM_USE_DEEP_GEMM=0
 
 ENV PYTHONPATH="/:/vllm-workspace"
 

README.md

@@ -8,7 +8,7 @@ Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https:
 
 ![vLLM worker banner](https://image.runpod.ai/preview/vllm/vllm-banner.png)
 
-Current vLLM version: [0.19.1](https://github.com/vllm-project/vllm/releases/tag/v0.19.1)
+Current vLLM version: [0.20.2](https://github.com/vllm-project/vllm/releases/tag/v0.20.2)
 
 
 > Check out our Load Balancer implementation here: [vLLM Load Balancer](https://github.com/runpod-workers/vllm-loadbalancer-ep)
@@ -47,7 +47,7 @@ Current vLLM version: [0.19.1](https://github.com/vllm-project/vllm/releases/tag
 **📦 Docker Image**: `runpod/worker-v1-vllm:<version>`
 
 - **Available Versions**: See [GitHub Releases](https://github.com/runpod-workers/worker-vllm/releases)
-- **CUDA Compatibility**: Requires CUDA >= 12.1
+- **CUDA Compatibility**: Requires CUDA >= 13.0
 
 ### Configuration
 

builder/requirements.txt

@@ -3,7 +3,7 @@ pandas
 pyarrow
 runpod==1.9.0
 huggingface-hub
-lmcache==0.4.2
+lmcache==0.4.5
 packaging>=24.2
 typing-extensions>=4.8.0
 pydantic

docs/configuration.md

@@ -97,10 +97,13 @@ 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`     | `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 — 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
 
 | Variable               | Default | Type/Choices                        | Description                                                                                       |

src/engine.py

@@ -1,4 +1,5 @@
 import asyncio
+import inspect
 import json
 import logging
 import os
@@ -7,6 +8,7 @@ from typing import AsyncGenerator, Optional
 
 from dotenv import load_dotenv
 from vllm import AsyncLLMEngine
+from vllm.inputs import TextPrompt
 from vllm.entrypoints.logger import RequestLogger
 from vllm.entrypoints.anthropic.protocol import AnthropicMessagesRequest, AnthropicMessagesResponse, AnthropicError, AnthropicErrorResponse
 from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
@@ -30,20 +32,33 @@ class vLLMEngine:
     def __init__(self, engine = None):
         load_dotenv() # For local development
         self.engine_args = get_engine_args()
-        logging.info(f"Engine args: {self.engine_args}")
 
-        # Initialize vLLM engine first
+        if engine is None:
-        self.llm = self._initialize_llm() if engine is None else engine.llm
+            ea = self.engine_args
+            summary = {
+                "model": ea.model,
+                "dtype": ea.dtype,
+                "quantization": ea.quantization,
+                "max_model_len": ea.max_model_len,
+                "tensor_parallel_size": ea.tensor_parallel_size,
+                "gpu_memory_utilization": ea.gpu_memory_utilization,
+            }
+            if ea.tokenizer and ea.tokenizer != ea.model:
+                summary["tokenizer"] = ea.tokenizer
+            logging.info("Engine config: %s", summary)
+            logging.debug("Full engine args: %s", ea)
 
-        # Only create custom tokenizer wrapper if not using mistral tokenizer mode
+            self.llm = self._initialize_llm()
-        # For mistral models, let vLLM handle tokenizer initialization
+
-        if self.engine_args.tokenizer_mode != 'mistral':
+            if self.engine_args.tokenizer_mode != 'mistral':
-            self.tokenizer = TokenizerWrapper(self.engine_args.tokenizer or self.engine_args.model, 
+                self.tokenizer = TokenizerWrapper(self.engine_args.tokenizer or self.engine_args.model,
-                                              self.engine_args.tokenizer_revision, 
+                                                  self.engine_args.tokenizer_revision,
-                                              self.engine_args.trust_remote_code)
+                                                  self.engine_args.trust_remote_code)
+            else:
+                self.tokenizer = None
         else:
-            # For mistral models, we'll get the tokenizer from vLLM later
+            self.llm = engine.llm
-            self.tokenizer = None
+            self.tokenizer = engine.tokenizer
             
         self.max_concurrency = int(os.getenv("MAX_CONCURRENCY", DEFAULT_MAX_CONCURRENCY))
         self.default_batch_size = int(os.getenv("DEFAULT_BATCH_SIZE", DEFAULT_BATCH_SIZE))
@@ -116,7 +131,7 @@ class vLLMEngine:
         if apply_chat_template or isinstance(llm_input, list):
             tokenizer_wrapper = self._get_tokenizer_for_chat_template()
             llm_input = tokenizer_wrapper.apply_chat_template(llm_input)
-        results_generator = self.llm.generate(llm_input, validated_sampling_params, request_id)
+        results_generator = self.llm.generate(TextPrompt(prompt=llm_input), validated_sampling_params, request_id)
         n_responses, n_input_tokens, is_first_output = validated_sampling_params.n, 0, True
         last_output_texts, token_counters = ["" for _ in range(n_responses)], {"batch": 0, "total": 0}
 
@@ -285,7 +300,6 @@ class OpenAIvLLMEngine(vLLMEngine):
         self.openai_serving_render = OpenAIServingRender(
             model_config=self.llm.model_config,
             renderer=self.llm.renderer,
-            io_processor=self.llm.io_processor,
             model_registry=self.serving_models.registry,
             request_logger=None,
             chat_template=chat_template,
@@ -357,8 +371,11 @@ class OpenAIvLLMEngine(vLLMEngine):
             enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
         )
 
-        if hasattr(self.chat_engine, 'warmup'):
+        warmup = getattr(self.chat_engine, 'warmup', None)
-            await self.chat_engine.warmup()
+        if callable(warmup):
+            result = warmup()
+            if inspect.isawaitable(result):
+                await result
 
     async def generate(self, openai_request: JobInput):
         # Ensure engines are ready (no-op if already initialized at startup)

src/tokenizer.py

@@ -1,10 +1,12 @@
-from transformers import AutoTokenizer
+import logging
 import os
 from typing import Union
 
+from transformers import AutoTokenizer
+
 class TokenizerWrapper:
     def __init__(self, tokenizer_name_or_path, tokenizer_revision, trust_remote_code):
-        print(f"tokenizer_name_or_path: {tokenizer_name_or_path}, tokenizer_revision: {tokenizer_revision}, trust_remote_code: {trust_remote_code}")
+        logging.debug("tokenizer_name_or_path: %s, tokenizer_revision: %s, trust_remote_code: %s", tokenizer_name_or_path, tokenizer_revision, trust_remote_code)
         self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, revision=tokenizer_revision or "main", trust_remote_code=trust_remote_code)
         self.custom_chat_template = os.getenv("CUSTOM_CHAT_TEMPLATE")
         self.has_chat_template = bool(self.tokenizer.chat_template) or bool(self.custom_chat_template)