#257 working vllm-0.4.0 pre PR version

vscentrum · Nov 7, 2024 · 87b5fd1 · 87b5fd1
1 parent 70084b2
commit 87b5fd1
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 8 deletions.
diff --git a/257_vLLM+ray/openai.eb b/257_vLLM+ray/openai.eb
@@ -19,7 +19,7 @@ builddependencies = [
 dependencies = [
     ('Python', '3.11.5'),
     ('SciPy-bundle', '2023.11'),
-    ('pydantic', '2.7.4'),
+    ('pydantic', '2.6.4'),
     ('typing-extensions', '4.10.0'),
     ('tqdm', '4.66.2'),
 ]

diff --git a/257_vLLM+ray/vllm-0.4.0.eb b/257_vLLM+ray/vllm-0.4.0.eb
@@ -16,7 +16,7 @@ toolchain = {'name': 'foss', 'version': '2023a'}
     # OK setuptools>=49.4.0
     # OK ninja  # For faster builds.
     # OK psutil
-    # OK    ray >= 2.9
+    # OK ray >= 2.9
     # OK sentencepiece  # Required for LLaMA tokenizer.
     # OK numpy
     # OK torch == 2.1.2
@@ -34,7 +34,7 @@ toolchain = {'name': 'foss', 'version': '2023a'}
     # OK tiktoken == 0.6.0 # Required for DBRX tokenizer
     # ok starlette 0.36.3 requires anyio, which is not installed.
     # NO sed out cmake and ninja from requirements.txt -> NO I can not - source is whl not sources.tar.gz
-        # OK -> move to deps? -> NO do not help -> sanity_pip_check = False -> OK
+        # OK -> move to deps? -> NO do not help -> sanity_pip_check = False -> OK -> hack from kenneth or use sources not whl
         # vllm 0.4.0 requires cmake, which is not installed.
         # vllm 0.4.0 requires ninja, which is not installed.
     # ok outlines 0.0.34 requires diskcache, which is not installed.
@@ -68,7 +68,7 @@ dependencies = [
     ('Triton', '2.1.0', versionsuffix),
     ('Ray-project', '2.9.1'),
     ('numba', '0.58.1'),
-    ('flash-attention', '2.6.3', versionsuffix)
+    # ('flash-attention', '2.6.3', versionsuffix),
 ]
 
 use_pip = True
@@ -125,16 +125,20 @@ exts_list = [
         'checksums': ['594e7204c770b47a62eb5c2ba7d25ea0ab2e16882b5f04556712a0228d3d3309'],
     }),
     (name, version, {
-        'source_tmpl': '%(namelower)s-%(version)s-cp311-cp311-manylinux1_x86_64.whl',
-        'checksums': ['69ce98bba1ad1c2174b4eb98502fb80181f1fcb7653551c512662da59cb18d8c'],
+        # delete cmake and ninja deps to let pip check pass
+        'preinstallopts': "sed -i -e '/cmake>=3.21/d' -e '/ninja/d' requirements.txt && ",
+        'source_urls': ['https://github.com/vllm-project/vllm/archive/'],
+        'sources': [{'download_filename': 'v%(version)s.tar.gz', 'filename': SOURCE_TAR_GZ}],
+        'checksums': ['95e5fa8f6ac63f11b4c7122e01c0892beae91f754a64a7e525bc06c3b49feac7'],
     }),
 ]
 
-sanity_pip_check = False
+sanity_pip_check = True
 
 moduleclass = 'ai'
 
-# E3:
+# E3: OK
+    # -> problem with prebuild whl -> use sources -> OK
     # add dep `auto_gptq`? (https://github.com/QwenLM/Qwen/issues/582)
     # another issue: https://github.com/vllm-project/vllm/issues/1344
     # NO -> missing dep flash-attention? (https://github.com/vllm-project/vllm/issues/8841) -> add flash-attention do deps -> not working
@@ -161,6 +165,7 @@ moduleclass = 'ai'
     # ImportError: /scratch/gent/vo/001/gvo00117/easybuild/RHEL8/cascadelake-ampere-ib/software/vLLM/0.4.0-foss-2023a-CUDA-12.1.1/lib/python3.11/site-packages/
     #     vllm/_C.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN3c106detail14torchCheckFailEPKcS2_jRKSs,  (at easybuild/easybuild-framework/easybuild/framework/easyblock.py:3670 in _sanity_check_step)
 # E2: OK
+    # -> another hack in meeting notes -> modify METADATA, but if I use source.tar.gz it is not necesary
     # -> add to deps does not work -> sanity_pip_check = False -> OK
     # vllm 0.4.0 requires cmake, which is not installed.
     # vllm 0.4.0 requires ninja, which is not installed.