feat: support phi3.5 moe (#2479)

* feat: support phi3.5 moe model loading

* fix: prefer llama base model and improve rotary logic

* feat: return reasonable generation and add integration test

* fix: run lint and update docs

* fix: rerun lint for openapi docs

* fix: prefer do_sample false unless temp is set by user, and update chat tests

* fix: small typo adjustments

* fix: consolidate long rope paths

* fix: revert greedy by default and test changes

* Vendor configuration so that we don't have to `trust_remote_code`

* Use SparseMoELayer

* Add support for dense MoE

* Some type annotations

* Add the usual model tests

* Ruff.

---------

Co-authored-by: Daniël de Kok <[email protected]>
Co-authored-by: Nicolas Patry <[email protected]>

docs/source/supported_models.md

@@ -20,6 +20,7 @@ Text Generation Inference enables serving optimized models on specific hardware
 - [Mixtral](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1)
 - [Gpt Bigcode](https://huggingface.co/bigcode/gpt_bigcode-santacoder)
 - [Phi](https://huggingface.co/microsoft/phi-1_5)
+- [PhiMoe](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct)
 - [Baichuan](https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat)
 - [Falcon](https://huggingface.co/tiiuae/falcon-7b-instruct)
 - [StarCoder 2](https://huggingface.co/bigcode/starcoder2-15b-instruct-v0.1)

integration-tests/models/__snapshots__/test_flash_phi35_moe/test_flash_phi35_moe.json

@@ -0,0 +1,109 @@
+{
+  "details": {
+    "best_of_sequences": null,
+    "finish_reason": "length",
+    "generated_tokens": 10,
+    "prefill": [
+      {
+        "id": 1724,
+        "logprob": null,
+        "text": "What"
+      },
+      {
+        "id": 338,
+        "logprob": -0.7133789,
+        "text": "is"
+      },
+      {
+        "id": 16030,
+        "logprob": -13.9296875,
+        "text": "gradient"
+      },
+      {
+        "id": 26815,
+        "logprob": -0.048919678,
+        "text": "descent"
+      },
+      {
+        "id": 29973,
+        "logprob": -3.0078125,
+        "text": "?"
+      },
+      {
+        "id": 13,
+        "logprob": -2.8105469,
+        "text": "\n"
+      },
+      {
+        "id": 13,
+        "logprob": -0.84521484,
+        "text": "\n"
+      }
+    ],
+    "seed": null,
+    "tokens": [
+      {
+        "id": 25584,
+        "logprob": -0.017028809,
+        "special": false,
+        "text": "Grad"
+      },
+      {
+        "id": 993,
+        "logprob": -0.0027313232,
+        "special": false,
+        "text": "ient"
+      },
+      {
+        "id": 26815,
+        "logprob": -0.023254395,
+        "special": false,
+        "text": " descent"
+      },
+      {
+        "id": 338,
+        "logprob": -2.0623207e-05,
+        "special": false,
+        "text": " is"
+      },
+      {
+        "id": 263,
+        "logprob": -0.5361328,
+        "special": false,
+        "text": " a"
+      },
+      {
+        "id": 937,
+        "logprob": -0.17578125,
+        "special": false,
+        "text": " first"
+      },
+      {
+        "id": 29899,
+        "logprob": 0.0,
+        "special": false,
+        "text": "-"
+      },
+      {
+        "id": 2098,
+        "logprob": -0.00011539459,
+        "special": false,
+        "text": "order"
+      },
+      {
+        "id": 13883,
+        "logprob": -0.47436523,
+        "special": false,
+        "text": " optimization"
+      },
+      {
+        "id": 5687,
+        "logprob": -0.00027680397,
+        "special": false,
+        "text": " algorithm"
+      }
+    ],
+    "top_tokens": null
+  },
+  "generated_text": "Gradient descent is a first-order optimization algorithm"
+}

integration-tests/models/__snapshots__/test_flash_phi35_moe/test_flash_phi35_moe_all_params.json

@@ -0,0 +1,99 @@
+{
+  "details": {
+    "best_of_sequences": null,
+    "finish_reason": "length",
+    "generated_tokens": 10,
+    "prefill": [
+      {
+        "id": 16030,
+        "logprob": null,
+        "text": "gradient"
+      },
+      {
+        "id": 26815,
+        "logprob": -6.4960938,
+        "text": "descent"
+      },
+      {
+        "id": 29973,
+        "logprob": -5.1484375,
+        "text": "?"
+      },
+      {
+        "id": 13,
+        "logprob": -4.0351562,
+        "text": "\n"
+      },
+      {
+        "id": 13,
+        "logprob": -5.2265625,
+        "text": "\n"
+      }
+    ],
+    "seed": 0,
+    "tokens": [
+      {
+        "id": 10994,
+        "logprob": -1.1542969,
+        "special": false,
+        "text": "Hello"
+      },
+      {
+        "id": 29991,
+        "logprob": 0.0,
+        "special": false,
+        "text": "!"
+      },
+      {
+        "id": 739,
+        "logprob": 0.0,
+        "special": false,
+        "text": " It"
+      },
+      {
+        "id": 2444,
+        "logprob": -0.42260742,
+        "special": false,
+        "text": " seems"
+      },
+      {
+        "id": 366,
+        "logprob": 0.0,
+        "special": false,
+        "text": " you"
+      },
+      {
+        "id": 29915,
+        "logprob": 0.0,
+        "special": false,
+        "text": "'"
+      },
+      {
+        "id": 276,
+        "logprob": -0.9838867,
+        "special": false,
+        "text": "re"
+      },
+      {
+        "id": 3211,
+        "logprob": 0.0,
+        "special": false,
+        "text": " address"
+      },
+      {
+        "id": 292,
+        "logprob": 0.0,
+        "special": false,
+        "text": "ing"
+      },
+      {
+        "id": 263,
+        "logprob": -0.15124512,
+        "special": false,
+        "text": " a"
+      }
+    ],
+    "top_tokens": null
+  },
+  "generated_text": "What is gradient descent?\n\nHello! It seems you're addressing a"
+}