Add initial support for compressed-tensors checkpoints

Dockerfile

@@ -247,7 +247,7 @@ COPY server/Makefile server/Makefile
 RUN cd server && \
     make gen-server && \
     pip install -r requirements_cuda.txt && \
-    pip install ".[bnb, accelerate, marlin, moe, quantize, peft, outlines]" --no-cache-dir && \
+    pip install ".[bnb, accelerate, compressed-tensors, marlin, moe, quantize, peft, outlines]" --no-cache-dir && \
     pip install nvidia-nccl-cu12==2.22.3
 
 ENV LD_PRELOAD=/opt/conda/lib/python3.11/site-packages/nvidia/nccl/lib/libnccl.so.2

Dockerfile_amd

@@ -296,7 +296,7 @@ COPY server/Makefile server/Makefile
 RUN cd server && \
     make gen-server && \
     pip install -r requirements_rocm.txt && \
-    pip install ".[accelerate, peft, outlines]" --no-cache-dir
+    pip install ".[accelerate, compressed-tensors, peft, outlines]" --no-cache-dir
 
 # Install benchmarker
 COPY --from=builder /usr/src/target/release-opt/text-generation-benchmark /usr/local/bin/text-generation-benchmark

Dockerfile_intel

@@ -102,7 +102,7 @@ COPY server/Makefile server/Makefile
 RUN cd server && \
     make gen-server && \
     pip install -r requirements_intel.txt && \
-    pip install ".[accelerate, peft, outlines]" --no-cache-dir
+    pip install ".[accelerate, compressed-tensors, peft, outlines]" --no-cache-dir
 
 ENV CCL_ROOT=/opt/intel/oneapi/ccl/latest
 ENV I_MPI_ROOT=/opt/intel/oneapi/mpi/latest

docs/source/reference/launcher.md

@@ -62,15 +62,16 @@ Options:
           [env: QUANTIZE=]
 
           Possible values:
-          - awq:              4 bit quantization. Requires a specific AWQ quantized model: <https://hf.co/models?search=awq>. Should replace GPTQ models wherever possible because of the better latency
-          - eetq:             8 bit quantization, doesn't require specific model. Should be a drop-in replacement to bitsandbytes with much better performance. Kernels are from <https://github.com/NetEase-FuXi/EETQ.git>
-          - exl2:             Variable bit quantization. Requires a specific EXL2 quantized model: <https://hf.co/models?search=exl2>. Requires exllama2 kernels and does not support tensor parallelism (num_shard > 1)
-          - gptq:             4 bit quantization. Requires a specific GTPQ quantized model: <https://hf.co/models?search=gptq>. text-generation-inference will use exllama (faster) kernels wherever possible, and use triton kernel (wider support) when it's not. AWQ has faster kernels
-          - marlin:           4 bit quantization. Requires a specific Marlin quantized model: <https://hf.co/models?search=marlin>
-          - bitsandbytes:     Bitsandbytes 8bit. Can be applied on any model, will cut the memory requirement in half, but it is known that the model will be much slower to run than the native f16
-          - bitsandbytes-nf4: Bitsandbytes 4bit. Can be applied on any model, will cut the memory requirement by 4x, but it is known that the model will be much slower to run than the native f16
-          - bitsandbytes-fp4: Bitsandbytes 4bit. nf4 should be preferred in most cases but maybe this one has better perplexity performance for you model
-          - fp8:              [FP8](https://developer.nvidia.com/blog/nvidia-arm-and-intel-publish-fp8-specification-for-standardization-as-an-interchange-format-for-ai/) (e4m3) works on H100 and above This dtype has native ops should be the fastest if available. This is currently not the fastest because of local unpacking + padding to satisfy matrix multiplication limitations
+          - awq:                4 bit quantization. Requires a specific AWQ quantized model: <https://hf.co/models?search=awq>. Should replace GPTQ models wherever possible because of the better latency
+          - compressed-tensors: Compressed tensors, which can be a mixture of different quantization methods
+          - eetq:               8 bit quantization, doesn't require specific model. Should be a drop-in replacement to bitsandbytes with much better performance. Kernels are from <https://github.com/NetEase-FuXi/EETQ.git>
+          - exl2:               Variable bit quantization. Requires a specific EXL2 quantized model: <https://hf.co/models?search=exl2>. Requires exllama2 kernels and does not support tensor parallelism (num_shard > 1)
+          - gptq:               4 bit quantization. Requires a specific GTPQ quantized model: <https://hf.co/models?search=gptq>. text-generation-inference will use exllama (faster) kernels wherever possible, and use triton kernel (wider support) when it's not. AWQ has faster kernels
+          - marlin:             4 bit quantization. Requires a specific Marlin quantized model: <https://hf.co/models?search=marlin>
+          - bitsandbytes:       Bitsandbytes 8bit. Can be applied on any model, will cut the memory requirement in half, but it is known that the model will be much slower to run than the native f16
+          - bitsandbytes-nf4:   Bitsandbytes 4bit. Can be applied on any model, will cut the memory requirement by 4x, but it is known that the model will be much slower to run than the native f16
+          - bitsandbytes-fp4:   Bitsandbytes 4bit. nf4 should be preferred in most cases but maybe this one has better perplexity performance for you model
+          - fp8:                [FP8](https://developer.nvidia.com/blog/nvidia-arm-and-intel-publish-fp8-specification-for-standardization-as-an-interchange-format-for-ai/) (e4m3) works on H100 and above This dtype has native ops should be the fastest if available. This is currently not the fastest because of local unpacking + padding to satisfy matrix multiplication limitations
 
 ```
 ## SPECULATE

flake.nix

@@ -5,7 +5,7 @@
       inputs.nixpkgs.follows = "tgi-nix/nixpkgs";
     };
     nix-filter.url = "github:numtide/nix-filter";
-    tgi-nix.url = "github:huggingface/text-generation-inference-nix";
+    tgi-nix.url = "github:huggingface/text-generation-inference-nix/compressed-tensors-0.7.1";
     nixpkgs.follows = "tgi-nix/nixpkgs";
     flake-utils.url = "github:numtide/flake-utils";
     rust-overlay = {

integration-tests/models/__snapshots__/test_compressed_tensors_w8an_fp/test_compressed_tensors_w8an.json

@@ -0,0 +1,104 @@
+{
+  "details": {
+    "best_of_sequences": null,
+    "finish_reason": "length",
+    "generated_tokens": 10,
+    "prefill": [
+      {
+        "id": 128000,
+        "logprob": null,
+        "text": "<|begin_of_text|>"
+      },
+      {
+        "id": 3923,
+        "logprob": -7.609375,
+        "text": "What"
+      },
+      {
+        "id": 374,
+        "logprob": -0.92529297,
+        "text": " is"
+      },
+      {
+        "id": 5655,
+        "logprob": -10.0,
+        "text": " deep"
+      },
+      {
+        "id": 6975,
+        "logprob": -0.94628906,
+        "text": " learning"
+      },
+      {
+        "id": 30,
+        "logprob": -2.9042969,
+        "text": "?"
+      }
+    ],
+    "seed": null,
+    "tokens": [
+      {
+        "id": 18682,
+        "logprob": -0.8769531,
+        "special": false,
+        "text": " Deep"
+      },
+      {
+        "id": 6975,
+        "logprob": -0.0076942444,
+        "special": false,
+        "text": " learning"
+      },
+      {
+        "id": 374,
+        "logprob": -0.25073242,
+        "special": false,
+        "text": " is"
+      },
+      {
+        "id": 264,
+        "logprob": -0.097595215,
+        "special": false,
+        "text": " a"
+      },
+      {
+        "id": 955,
+        "logprob": -0.921875,
+        "special": false,
+        "text": " type"
+      },
+      {
+        "id": 315,
+        "logprob": -0.00027918816,
+        "special": false,
+        "text": " of"
+      },
+      {
+        "id": 21075,
+        "logprob": -0.5527344,
+        "special": false,
+        "text": " artificial"
+      },
+      {
+        "id": 11478,
+        "logprob": -0.042541504,
+        "special": false,
+        "text": " intelligence"
+      },
+      {
+        "id": 320,
+        "logprob": -0.38891602,
+        "special": false,
+        "text": " ("
+      },
+      {
+        "id": 15836,
+        "logprob": -0.0011043549,
+        "special": false,
+        "text": "AI"
+      }
+    ],
+    "top_tokens": null
+  },
+  "generated_text": " Deep learning is a type of artificial intelligence (AI"
+}

integration-tests/models/__snapshots__/test_compressed_tensors_w8an_fp/test_compressed_tensors_w8an_all_params.json

@@ -0,0 +1,99 @@
+{
+  "details": {
+    "best_of_sequences": null,
+    "finish_reason": "length",
+    "generated_tokens": 10,
+    "prefill": [
+      {
+        "id": 128000,
+        "logprob": null,
+        "text": "<|begin_of_text|>"
+      },
+      {
+        "id": 3923,
+        "logprob": -7.609375,
+        "text": "What"
+      },
+      {
+        "id": 374,
+        "logprob": -0.92529297,
+        "text": " is"
+      },
+      {
+        "id": 5655,
+        "logprob": -10.0,
+        "text": " deep"
+      },
+      {
+        "id": 6975,
+        "logprob": -0.94628906,
+        "text": " learning"
+      }
+    ],
+    "seed": 0,
+    "tokens": [
+      {
+        "id": 5380,
+        "logprob": -0.23840332,
+        "special": false,
+        "text": "?\n"
+      },
+      {
+        "id": 34564,
+        "logprob": 0.0,
+        "special": false,
+        "text": "Deep"
+      },
+      {
+        "id": 6975,
+        "logprob": 0.0,
+        "special": false,
+        "text": " learning"
+      },
+      {
+        "id": 11,
+        "logprob": 0.0,
+        "special": false,
+        "text": ","
+      },
+      {
+        "id": 1101,
+        "logprob": -1.2011719,
+        "special": false,
+        "text": " also"
+      },
+      {
+        "id": 3967,
+        "logprob": 0.0,
+        "special": false,
+        "text": " known"
+      },
+      {
+        "id": 439,
+        "logprob": 0.0,
+        "special": false,
+        "text": " as"
+      },
+      {
+        "id": 30828,
+        "logprob": 0.0,
+        "special": false,
+        "text": " neural"
+      },
+      {
+        "id": 4009,
+        "logprob": -0.6777344,
+        "special": false,
+        "text": " network"
+      },
+      {
+        "id": 477,
+        "logprob": 0.0,
+        "special": false,
+        "text": " or"
+      }
+    ],
+    "top_tokens": null
+  },
+  "generated_text": "What is deep learning?\nDeep learning, also known as neural network or"
+}