···

       
7
       
7
       
 
       
  fetchFromGitHub,

     

       
8
       
8
       
 
       
  symlinkJoin,

     

       
9
       
9
       
 
       
  autoAddDriverRunpath,

     

       
10
       
10
       
-
       
  fetchpatch2,

     

       
11
       
10
       
 
       


     

       
12
       
11
       
 
       
  # build system

     

       
13
       
12
       
 
       
  cmake,

     
···

       
71
       
70
       
 
       
  bitsandbytes,

     

       
72
       
71
       
 
       
  flashinfer,

     

       
73
       
72
       
 
       
  py-libnuma,

     

       
73
       
73
       
+
       
  setproctitle,

     

       
74
       
74
       
+
       
  openai-harmony,

     

       
74
       
75
       
 
       


     

       
75
       
76
       
 
       
  # internal dependency - for overriding in overlays

     

       
76
       
77
       
 
       
  vllm-flash-attn ? null,

     
···

       
94
       
95
       
 
       
  shouldUsePkg =

     

       
95
       
96
       
 
       
    pkg: if pkg != null && lib.meta.availableOn stdenv.hostPlatform pkg then pkg else null;

     

       
96
       
97
       
 
       


     

       
97
       
97
       
-
       
  # see CMakeLists.txt, grepping for GIT_TAG near cutlass

     

       
98
       
98
       
+
       
  # see CMakeLists.txt, grepping for CUTLASS_REVISION

     

       
98
       
99
       
 
       
  # https://github.com/vllm-project/vllm/blob/v${version}/CMakeLists.txt

     

       
99
       
100
       
 
       
  cutlass = fetchFromGitHub {

     

       
100
       
101
       
 
       
    owner = "NVIDIA";

     
···

       
113
       
114
       
 
       
    src = fetchFromGitHub {

     

       
114
       
115
       
 
       
      owner = "vllm-project";

     

       
115
       
116
       
 
       
      repo = "FlashMLA";

     

       
116
       
116
       
-
       
      rev = "575f7724b9762f265bbee5889df9c7d630801845";

     

       
117
       
117
       
-
       
      hash = "sha256-8WrKMl0olr0nYV4FRJfwSaJ0F5gWQpssoFMjr9tbHBk=";

     

       
117
       
117
       
+
       
      rev = "0e43e774597682284358ff2c54530757b654b8d1";

     

       
118
       
118
       
+
       
      hash = "sha256-wxL/jtq/lsLg1o+4392KNgfw5TYlW6lqEVbmR3Jl4/Q=";

     

       
118
       
119
       
 
       
    };

     

       
119
       
120
       
 
       


     

       
120
       
121
       
 
       
    dontConfigure = true;

     
···

       
140
       
141
       
 
       
    src = fetchFromGitHub {

     

       
141
       
142
       
 
       
      owner = "vllm-project";

     

       
142
       
143
       
 
       
      repo = "flash-attention";

     

       
143
       
143
       
-
       
      rev = "1c2624e53c078854e0637ee566c72fe2107e75f4";

     

       
144
       
144
       
-
       
      hash = "sha256-WWFhHEUSAlsXr2yR4rGlTQQnSafXKg8gO5PQA8HPYGE=";

     

       
144
       
144
       
+
       
      rev = "57b4e68b9f9d94750b46de8f8dbd2bfcc86edd4f";

     

       
145
       
145
       
+
       
      hash = "sha256-c7L7WZVVEnXMOTPBoSp7jhkl9d4TA4sj11QvOSWTDIE=";

     

       
145
       
146
       
 
       
    };

     

       
146
       
147
       
 
       


     

       
147
       
148
       
 
       
    dontConfigure = true;

     
···

       
233
       
234
       
 
       
    libcusolver # cusolverDn.h

     

       
234
       
235
       
 
       
    cuda_nvtx

     

       
235
       
236
       
 
       
    cuda_nvrtc

     

       
237
       
237
       
+
       
    # cusparselt # cusparseLt.h

     

       
236
       
238
       
 
       
    libcublas

     

       
237
       
239
       
 
       
  ];

     

       
238
       
240
       
 
       


     
···

       
249
       
251
       
 
       


     

       
250
       
252
       
 
       
buildPythonPackage rec {

     

       
251
       
253
       
 
       
  pname = "vllm";

     

       
252
       
252
       
-
       
  version = "0.10.0";

     

       
254
       
254
       
+
       
  version = "0.10.1.1";

     

       
253
       
255
       
 
       
  pyproject = true;

     

       
254
       
256
       
 
       


     

       
255
       
257
       
 
       
  # https://github.com/vllm-project/vllm/issues/12083

     
···

       
261
       
263
       
 
       
    owner = "vllm-project";

     

       
262
       
264
       
 
       
    repo = "vllm";

     

       
263
       
265
       
 
       
    tag = "v${version}";

     

       
264
       
264
       
-
       
    hash = "sha256-R9arpFz+wkDGmB3lW+H8d/37EoAQDyCWjLHJW1VTutk=";

     

       
266
       
266
       
+
       
    hash = "sha256-lLNjBv5baER0AArX3IV4HWjDZ2jTGXyGIvnHupR8MGM=";

     

       
265
       
267
       
 
       
  };

     

       
266
       
268
       
 
       


     

       
267
       
269
       
 
       
  patches = [

     

       
268
       
268
       
-
       
    # error: ‘BF16Vec16’ in namespace ‘vec_op’ does not name a type; did you mean ‘FP16Vec16’?

     

       
269
       
269
       
-
       
    # Reported: https://github.com/vllm-project/vllm/issues/21714

     

       
270
       
270
       
-
       
    # Fix from https://github.com/vllm-project/vllm/pull/21848

     

       
271
       
271
       
-
       
    (fetchpatch2 {

     

       
272
       
272
       
-
       
      name = "build-fix-for-arm-without-bf16";

     

       
273
       
273
       
-
       
      url = "https://github.com/vllm-project/vllm/commit/b876860c6214d03279e79e0babb7eb4e3e286cbd.patch";

     

       
274
       
274
       
-
       
      hash = "sha256-tdBAObFxliVUNTWeSggaLtS4K9f8zEVu22nSgRmMsDs=";

     

       
275
       
275
       
-
       
    })

     

       
276
       
270
       
 
       
    ./0002-setup.py-nix-support-respect-cmakeFlags.patch

     

       
277
       
271
       
 
       
    ./0003-propagate-pythonpath.patch

     

       
278
       
272
       
 
       
    ./0005-drop-intel-reqs.patch

     
···

       
394
       
388
       
 
       
    opentelemetry-api

     

       
395
       
389
       
 
       
    opentelemetry-exporter-otlp

     

       
396
       
390
       
 
       
    bitsandbytes

     

       
391
       
391
       
+
       
    setproctitle

     

       
392
       
392
       
+
       
    openai-harmony

     

       
397
       
393
       
 
       
    # vLLM needs Torch's compiler to be present in order to use torch.compile

     

       
398
       
394
       
 
       
    torch.stdenv.cc

     

       
399
       
395
       
 
       
  ]