Test Marlin MoE with desc_act=true (#2622)

Update the Mixtral GPTQ test to use a model with `desc_act=true` and `group_size!=-1` to ensure that we are checking activation sorting/non-full K (with tensor parallelism). The `desc_act=false` case is already checked by the Mixtral AWQ test.
huggingface · Oct 21, 2024 · 7f54b73 · 7f54b73
1 parent 5e0fb46
commit 7f54b73
Show file tree

Hide file tree

Showing 4 changed files with 301 additions and 199 deletions.
diff --git a/integration-tests/models/__snapshots__/test_flash_mixtral_gptq/test_flash_mixtral_gptq.json b/integration-tests/models/__snapshots__/test_flash_mixtral_gptq/test_flash_mixtral_gptq.json
@@ -10,80 +10,95 @@
         "text": "<s>"
       },
       {
-        "id": 3735,
-        "logprob": -11.0078125,
-        "text": "Test"
+        "id": 1824,
+        "logprob": -9.2890625,
+        "text": "What"
       },
       {
-        "id": 2159,
-        "logprob": -13.59375,
-        "text": "request"
+        "id": 349,
+        "logprob": -1.1503906,
+        "text": "is"
+      },
+      {
+        "id": 3534,
+        "logprob": -9.5859375,
+        "text": "deep"
+      },
+      {
+        "id": 5168,
+        "logprob": -1.3945312,
+        "text": "learning"
+      },
+      {
+        "id": 28804,
+        "logprob": -0.4555664,
+        "text": "?"
       }
     ],
     "seed": null,
     "tokens": [
       {
         "id": 13,
-        "logprob": -1.7089844,
+        "logprob": -0.6953125,
         "special": false,
         "text": "\n"
       },
       {
         "id": 13,
-        "logprob": -0.68847656,
+        "logprob": -0.4777832,
         "special": false,
         "text": "\n"
       },
       {
-        "id": 28771,
-        "logprob": -1.9394531,
+        "id": 23229,
+        "logprob": -0.13256836,
         "special": false,
-        "text": "#"
+        "text": "Deep"
       },
       {
-        "id": 3735,
-        "logprob": -2.8808594,
+        "id": 5168,
+        "logprob": -0.023849487,
         "special": false,
-        "text": " Test"
+        "text": " learning"
       },
       {
-        "id": 2159,
-        "logprob": -0.37280273,
+        "id": 349,
+        "logprob": -0.13977051,
         "special": false,
-        "text": " request"
+        "text": " is"
       },
       {
-        "id": 13,
-        "logprob": -0.26098633,
+        "id": 264,
+        "logprob": -0.14489746,
         "special": false,
-        "text": "\n"
+        "text": " a"
       },
       {
-        "id": 13,
-        "logprob": -0.0017137527,
+        "id": 19804,
+        "logprob": -0.63183594,
         "special": false,
-        "text": "\n"
+        "text": " subset"
       },
       {
-        "id": 1064,
-        "logprob": -2.2695312,
+        "id": 302,
+        "logprob": -0.010314941,
         "special": false,
-        "text": "##"
+        "text": " of"
       },
       {
-        "id": 3735,
-        "logprob": -1.9238281,
+        "id": 5599,
+        "logprob": -0.0635376,
         "special": false,
-        "text": " Test"
+        "text": " machine"
       },
       {
-        "id": 2159,
-        "logprob": -0.48828125,
+        "id": 5168,
+        "logprob": -0.0028572083,
         "special": false,
-        "text": " request"
+        "text": " learning"
       }
     ],
     "top_tokens": null
   },
-  "generated_text": "\n\n# Test request\n\n## Test request"
+  "generated_text": "\n\nDeep learning is a subset of machine learning"
 }
diff --git a/...ests/models/__snapshots__/test_flash_mixtral_gptq/test_flash_mixtral_gptq_all_params.json b/...ests/models/__snapshots__/test_flash_mixtral_gptq/test_flash_mixtral_gptq_all_params.json
@@ -10,80 +10,90 @@
         "text": "<s>"
       },
       {
-        "id": 3735,
-        "logprob": -11.0078125,
-        "text": "Test"
+        "id": 349,
+        "logprob": -12.0546875,
+        "text": "is"
       },
       {
-        "id": 2159,
-        "logprob": -13.59375,
-        "text": "request"
+        "id": 3534,
+        "logprob": -10.53125,
+        "text": "deep"
+      },
+      {
+        "id": 5168,
+        "logprob": -2.71875,
+        "text": "learning"
+      },
+      {
+        "id": 28804,
+        "logprob": -5.0078125,
+        "text": "?"
       }
     ],
     "seed": 0,
     "tokens": [
       {
         "id": 13,
-        "logprob": -0.34838867,
+        "logprob": 0.0,
         "special": false,
         "text": "\n"
       },
       {
-        "id": 13940,
-        "logprob": -0.38916016,
+        "id": 23229,
+        "logprob": -0.18237305,
         "special": false,
-        "text": "``"
+        "text": "Deep"
       },
       {
-        "id": 28832,
+        "id": 17504,
         "logprob": 0.0,
         "special": false,
-        "text": "`"
+        "text": " Learning"
       },
       {
-        "id": 3371,
-        "logprob": -1.2529297,
+        "id": 349,
+        "logprob": 0.0,
         "special": false,
-        "text": "json"
+        "text": " is"
       },
       {
-        "id": 13,
+        "id": 264,
         "logprob": 0.0,
         "special": false,
-        "text": "\n"
+        "text": " a"
       },
       {
-        "id": 28751,
+        "id": 19804,
         "logprob": 0.0,
         "special": false,
-        "text": "{"
+        "text": " subset"
       },
       {
-        "id": 13,
+        "id": 302,
         "logprob": 0.0,
         "special": false,
-        "text": "\n"
+        "text": " of"
       },
       {
-        "id": 2287,
-        "logprob": 0.0,
+        "id": 13253,
+        "logprob": -0.6040039,
         "special": false,
-        "text": "   "
+        "text": " Machine"
       },
       {
-        "id": 345,
+        "id": 17504,
         "logprob": 0.0,
         "special": false,
-        "text": " \""
+        "text": " Learning"
       },
       {
-        "id": 3134,
-        "logprob": -0.640625,
+        "id": 28725,
+        "logprob": -0.11621094,
         "special": false,
-        "text": "request"
+        "text": ","
       }
     ],
     "top_tokens": null
   },
-  "generated_text": "Test request\n```json\n{\n    \"request"
+  "generated_text": "What is deep learning?\nDeep Learning is a subset of Machine Learning,"
 }