Tmp tp transformers (#2942)

* Upgrade the version number. * Remove modifications in Lock. * Tmp branch to test transformers backend with 2.5.1 and TP>1 * Fixing the transformers backend. inference_mode forces the use of `aten.matmul` instead of `aten.mm` the former doesn't have sharding support crashing the transformers TP support. `lm_head.forward` also crashes because it skips the hook that cast/decast the DTensor. Torch 2.5.1 is required for sharding support. * Put back the attention impl. * Revert the flashinfer (this will fails). * Building AOT. * Using 2.5 kernels. * Remove the archlist, it's defined in the docker anyway.
2025-04-19 22:02:06 +00:00 · 2025-01-23 18:07:30 +01:00 · 2025-01-23 18:07:30 +01:00 · 29a0893b67
commit 29a0893b67
parent 0a89902663
15 changed files with 47 additions and 49 deletions
--- a/6
+++ b/6
@ -47,7 +47,7 @@ RUN cargo build --profile release-opt --frozen
 FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS pytorch-install

 # NOTE: When updating PyTorch version, beware to remove `pip install nvidia-nccl-cu12==2.22.3` below in the Dockerfile. Context: https://github.com/huggingface/text-generation-inference/pull/2099
-ARG PYTORCH_VERSION=2.4.0
+ARG PYTORCH_VERSION=2.5.1

 ARG PYTHON_VERSION=3.11
 # Keep in sync with `server/pyproject.toml
@ -235,8 +235,8 @@ RUN cd server && \
    make gen-server && \
    python -c "from text_generation_server.pb import generate_pb2" && \
    pip install -U pip uv && \
-    uv pip install -e ".[attention, bnb, accelerate, compressed-tensors, marlin, moe, quantize, peft, outlines]" --no-cache-dir && \
-    uv pip install nvidia-nccl-cu12==2.22.3
+    uv pip install -e ".[attention, bnb, accelerate, compressed-tensors, marlin, moe, quantize, peft, outlines]" --no-cache-dir # && \
+    # uv pip install nvidia-nccl-cu12==2.22.3

 ENV LD_PRELOAD=/opt/conda/lib/python3.11/site-packages/nvidia/nccl/lib/libnccl.so.2
 # Required to find libpython within the rust binaries
--- a/README.md
+++ b/README.md
@ -84,7 +84,7 @@ model=HuggingFaceH4/zephyr-7b-beta
 volume=$PWD/data

 docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.0 --model-id $model
+    ghcr.io/huggingface/text-generation-inference:3.0.2 --model-id $model
 ```

 And then you can make requests like
@ -121,7 +121,7 @@ curl localhost:8080/v1/chat/completions \

 **Note:** To use NVIDIA GPUs, you need to install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html). We also recommend using NVIDIA drivers with CUDA version 12.2 or higher. For running the Docker container on a machine with no GPUs or CUDA support, it is enough to remove the `--gpus all` flag and add `--disable-custom-kernels`, please note CPU is not the intended platform for this project, so performance might be subpar.

-**Note:** TGI supports AMD Instinct MI210 and MI250 GPUs. Details can be found in the [Supported Hardware documentation](https://huggingface.co/docs/text-generation-inference/installation_amd#using-tgi-with-amd-gpus). To use AMD GPUs, please use `docker run --device /dev/kfd --device /dev/dri --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.0-rocm --model-id $model` instead of the command above.
+**Note:** TGI supports AMD Instinct MI210 and MI250 GPUs. Details can be found in the [Supported Hardware documentation](https://huggingface.co/docs/text-generation-inference/installation_amd#using-tgi-with-amd-gpus). To use AMD GPUs, please use `docker run --device /dev/kfd --device /dev/dri --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.2-rocm --model-id $model` instead of the command above.

 To see all options to serve your models (in the [code](https://github.com/huggingface/text-generation-inference/blob/main/launcher/src/main.rs) or in the cli):
 ```
@ -152,7 +152,7 @@ volume=$PWD/data # share a volume with the Docker container to avoid downloading
 token=<your cli READ token>

 docker run --gpus all --shm-size 1g -e HF_TOKEN=$token -p 8080:80 -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.0 --model-id $model
+    ghcr.io/huggingface/text-generation-inference:3.0.2 --model-id $model
 ```

 ### A note on Shared Memory (shm)
--- a/docs/source/basic_tutorials/gated_model_access.md
+++ b/docs/source/basic_tutorials/gated_model_access.md
@ -19,6 +19,6 @@ docker run --gpus all \
    --shm-size 1g \
    -e HF_TOKEN=$token \
    -p 8080:80 \
-    -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.1 \
+    -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.2 \
    --model-id $model
 ```
--- a/docs/source/conceptual/quantization.md
+++ b/docs/source/conceptual/quantization.md
@ -19,7 +19,7 @@ bitsandbytes is a library used to apply 8-bit and 4-bit quantization to models.
 In TGI, you can use 8-bit quantization by adding `--quantize bitsandbytes` like below 👇

 ```bash
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.1 --model-id $model --quantize bitsandbytes
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.2 --model-id $model --quantize bitsandbytes
 ```

 4-bit quantization is also possible with bitsandbytes. You can choose one of the following 4-bit data types: 4-bit float (`fp4`), or 4-bit `NormalFloat` (`nf4`). These data types were introduced in the context of parameter-efficient fine-tuning, but you can apply them for inference by automatically converting the model weights on load.
@ -27,7 +27,7 @@ docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingf
 In TGI, you can use 4-bit quantization by adding `--quantize bitsandbytes-nf4` or `--quantize bitsandbytes-fp4` like below 👇

 ```bash
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.1 --model-id $model --quantize bitsandbytes-nf4
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.2 --model-id $model --quantize bitsandbytes-nf4
 ```

 You can get more information about 8-bit quantization by reading this [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration), and 4-bit quantization by reading [this blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes).
@ -48,7 +48,7 @@ $$({\hat{W}_{l}}^{*} = argmin_{\hat{W_{l}}} ||W_{l}X-\hat{W}_{l}X||^{2}_{2})$$
 TGI allows you to both run an already GPTQ quantized model (see available models [here](https://huggingface.co/models?search=gptq)) or quantize a model of your choice using quantization script. You can run a quantized model by simply passing --quantize like below 👇

 ```bash
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.1 --model-id $model --quantize gptq
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:3.0.2 --model-id $model --quantize gptq
 ```

 Note that TGI's GPTQ implementation doesn't use [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ) under the hood. However, models quantized using AutoGPTQ or Optimum can still be served by TGI.
--- a/docs/source/installation_amd.md
+++ b/docs/source/installation_amd.md
@ -11,7 +11,7 @@ volume=$PWD/data # share a volume with the Docker container to avoid downloading
 docker run --rm -it --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
    --device=/dev/kfd --device=/dev/dri --group-add video \
    --ipc=host --shm-size 256g --net host -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.1-rocm \
+    ghcr.io/huggingface/text-generation-inference:3.0.2-rocm \
    --model-id $model
 ```

--- a/docs/source/installation_intel.md
+++ b/docs/source/installation_intel.md
@ -12,7 +12,7 @@ volume=$PWD/data # share a volume with the Docker container to avoid downloading
 docker run --rm --privileged --cap-add=sys_nice \
    --device=/dev/dri \
    --ipc=host --shm-size 1g --net host -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.1-intel-xpu \
+    ghcr.io/huggingface/text-generation-inference:3.0.2-intel-xpu \
    --model-id $model --cuda-graphs 0
 ```

@ -29,7 +29,7 @@ volume=$PWD/data # share a volume with the Docker container to avoid downloading
 docker run --rm --privileged --cap-add=sys_nice \
    --device=/dev/dri \
    --ipc=host --shm-size 1g --net host -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.1-intel-cpu \
+    ghcr.io/huggingface/text-generation-inference:3.0.2-intel-cpu \
    --model-id $model --cuda-graphs 0
 ```

--- a/docs/source/installation_nvidia.md
+++ b/docs/source/installation_nvidia.md
@ -11,7 +11,7 @@ model=teknium/OpenHermes-2.5-Mistral-7B
 volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

 docker run --gpus all --shm-size 64g -p 8080:80 -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.1 \
+    ghcr.io/huggingface/text-generation-inference:3.0.2 \
    --model-id $model
 ```

--- a/docs/source/quicktour.md
+++ b/docs/source/quicktour.md
@ -11,7 +11,7 @@ model=teknium/OpenHermes-2.5-Mistral-7B
 volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

 docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data \
-    ghcr.io/huggingface/text-generation-inference:3.0.1 \
+    ghcr.io/huggingface/text-generation-inference:3.0.2 \
    --model-id $model
 ```

@ -96,7 +96,7 @@ curl 127.0.0.1:8080/generate \
 To see all possible deploy flags and options, you can use the `--help` flag. It's possible to configure the number of shards, quantization, generation parameters, and more.

 ```bash
-docker run ghcr.io/huggingface/text-generation-inference:3.0.1 --help
+docker run ghcr.io/huggingface/text-generation-inference:3.0.2 --help
 ```

 </Tip>
--- a/docs/source/reference/api_reference.md
+++ b/docs/source/reference/api_reference.md
@ -163,7 +163,7 @@ hub = {

 # create Hugging Face Model Class
 huggingface_model = HuggingFaceModel(
- image_uri=get_huggingface_llm_image_uri("huggingface",version="3.0.1"),
+ image_uri=get_huggingface_llm_image_uri("huggingface",version="3.0.2"),
 env=hub,
 role=role,
 )
--- a/server/Makefile-flashinfer
+++ b/server/Makefile-flashinfer
@ -1,5 +1,6 @@
 install-flashinfer:
 	# We need fsspec as an additional dependency, but
 	# `pip install flashinfer` cannot resolve it.
-	pip install fsspec
-	pip install flashinfer==0.2.0.post1 -i https://flashinfer.ai/whl/cu124/torch2.4
+	pip install fsspec sympy==1.13.1 numpy
+	pip install -U setuptools
+	FLASHINFER_ENABLE_AOT=1 pip install git+https://github.com/flashinfer-ai/flashinfer.git@v0.2.0.post1#egg=flashinfer  --no-build-isolation
--- a/server/pyproject.toml
+++ b/server/pyproject.toml
@ -69,22 +69,22 @@ gen = [

 [tool.uv.sources]
 attention-kernels = [
-  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.4-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
-  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.4-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
-  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.4-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
-  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.4-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
+  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.5-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
+  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.5-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
+  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.5-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
+  { url = "https://github.com/danieldk/attention-kernels/releases/download/v0.1.1/attention_kernels-0.1.1+cu123torch2.5-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
 ]
 marlin-kernels = [
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.4-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.4-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.4-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.4-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.5-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.5-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.5-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.7/marlin_kernels-0.3.7+cu123torch2.5-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
 ]
 moe-kernels = [
-  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.4-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
-  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.4-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
-  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.4-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
-  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.4-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
+  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.5-cp39-cp39-linux_x86_64.whl", marker = "python_version == '3.9'" },
+  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.5-cp310-cp310-linux_x86_64.whl", marker = "python_version == '3.10'" },
+  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.5-cp311-cp311-linux_x86_64.whl", marker = "python_version == '3.11'" },
+  { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.7.0/moe_kernels-0.7.0+cu123torch2.5-cp312-cp312-linux_x86_64.whl", marker = "python_version == '3.12'" },
 ]

 [tool.pytest.ini_options]
--- a/server/text_generation_server/models/init.py
+++ b/server/text_generation_server/models/init.py
@ -931,10 +931,10 @@ def get_model(
                trust_remote_code=trust_remote_code,
                lora_adapter_ids=lora_adapter_ids,
            )
-        elif sharded:
-            raise NotImplementedError(
-                FLASH_ATT_ERROR_MESSAGE.format(f"Sharded {model_type}")
-            )
+        # elif sharded:
+        #     raise NotImplementedError(
+        #         FLASH_ATT_ERROR_MESSAGE.format(f"Sharded {model_type}")
+        #     )
        else:
            return transformers_causal_lm_class.fallback(
                model_id,
--- a/server/text_generation_server/models/transformers_flash_causal_lm.py
+++ b/server/text_generation_server/models/transformers_flash_causal_lm.py
@ -3,7 +3,7 @@ from typing import List, Optional

 import torch
 from opentelemetry import trace
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+from transformers import AutoTokenizer, AutoModelForCausalLM
 import transformers.modeling_utils

 from text_generation_server.models.flash_causal_lm import FlashCausalLM
@ -36,9 +36,7 @@ def tgi_flash_attention_forward(
    softcap: Optional[float] = None,
    **kwargs,  # This is needed to "absorb" other args passed by Transformers modeling
 ):
-
    kv_cache = kv_cache[module.layer_idx]
-
    query_states = query_states.transpose(1, 2).squeeze(dim=0)
    key_states = key_states.transpose(1, 2).squeeze(dim=0)
    value_states = value_states.transpose(1, 2).squeeze(dim=0)
@ -95,7 +93,6 @@ class TransformersFlashCausalLM(FlashCausalLM):
        default_dtype=torch.float16,
        trust_remote_code: bool = False,
        tokenizer_class=AutoTokenizer,
-        config_class=AutoConfig,
        kv_cache_dtype: Optional[torch.dtype] = None,
    ):
        self.quantize = quantize
@ -105,17 +102,17 @@ class TransformersFlashCausalLM(FlashCausalLM):
            raise RuntimeError("Speculator decoding is not enabled for AutoModel")

        if torch.cuda.is_available():
-            device = torch.device("cuda:0")
-            dtype = torch.float16 if dtype is None else dtype
+            device = torch.device(f"cuda:{rank}")
+            dtype = default_dtype if dtype is None else dtype
        elif hasattr(torch, "xpu") and torch.xpu.is_available():
            device = torch.device("xpu")
-            dtype = torch.float16 if dtype is None else dtype
+            dtype = default_dtype if dtype is None else dtype
        else:
            raise ValueError(
                "Flash `Transformers` modeling backend is not available on cpu."
            )

-        tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer = tokenizer_class.from_pretrained(
            model_id,
            revision=revision,
            padding_side="left",
@ -126,10 +123,10 @@ class TransformersFlashCausalLM(FlashCausalLM):
            model_id,
            revision=revision,
            torch_dtype=dtype,
-            device_map="auto",
            load_in_8bit=quantize == "bitsandbytes",
            trust_remote_code=trust_remote_code,
            attn_implementation="tgi",
+            device_map=device if world_size == 1 else None,
            tp_plan="auto" if world_size > 1 else None,
        )

@ -261,6 +258,6 @@ class TransformersFlashCausalLM(FlashCausalLM):
        # To update with full Transformers support asap
        if lm_head_indices is not None:
            hidden_states = hidden_states[lm_head_indices]
-        logits = self.model.lm_head.forward(hidden_states)
+        logits = self.model.lm_head(hidden_states)

        return logits, None
--- a/server/text_generation_server/server.py
+++ b/server/text_generation_server/server.py
@ -68,9 +68,9 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
        self.quantize = model.quantize
        self.server_urls = server_urls
        # For some reason, inference_mode does not work well with GLOO which we use on CPU
-        if model.device.type == "cuda":
-            # Force inference mode for the lifetime of TextGenerationService
-            self._inference_mode_raii_guard = torch._C._InferenceMode(True)
+        # if model.device.type == "cuda":
+        #     # Force inference mode for the lifetime of TextGenerationService
+        #     self._inference_mode_raii_guard = torch._C._InferenceMode(True)

    async def Info(self, request, context):
        return self.model.info