[CICD] Unify Qwen3.6 tests and consolidate MUSA/MetaX/Ascend CI updates - #347
Conversation
aba4d82 to
41f0839
Compare
fec105e to
db5055a
Compare
HermiaHuan
changed the title
Unify Qwen3.6 model test configs across platforms and consolidate pending MUSA S5000, MetaX graph serving, and Ascend 910C CI updates.
db5055a to
03b71a1
Compare
| platform: hygon | ||
| secrets: inherit | ||
|
|
||
| # ============================================================ |
There was a problem hiding this comment.
why we change the model path in this file
| generate: | ||
| prompts: | ||
| - "Introduce yourself,please" | ||
| - "Where is the capital of France" |
There was a problem hiding this comment.
need to validate the output response
| tensor_parallel_size: 2 | ||
| pipeline_parallel_size: 1 | ||
| max_model_len: 262144 | ||
| max_model_len: 8192 |
There was a problem hiding this comment.
Is this the correct setting?
| ``FL_TEST_PLATFORM`` when unset. | ||
| device: Optional device name used to apply device-level overrides. | ||
| Defaults to ``FL_TEST_DEVICE`` when unset. | ||
| platforms_dir: Override directory for platform configs. |
There was a problem hiding this comment.
why we need to change this core file
There was a problem hiding this comment.
need to remove this workflows
There was a problem hiding this comment.
need to modify the configure file.
| chat_messages: | ||
| - role: "user" | ||
| content: "Introduce yourself,please" | ||
| content: "Where is the capital of France" |
There was a problem hiding this comment.
chat_messages should be removed.
| max_tokens: 1024 | ||
| sampling: | ||
| temperature: 0.0 | ||
| stream: false |
There was a problem hiding this comment.
why we change stream to false?
| sampling: | ||
| temperature: 0.0 | ||
| stream: false | ||
| max_tokens: 256 |
| max_tokens: 256 | ||
| chat_messages: | ||
| - role: "user" | ||
| content: "Where is the capital of France" |
There was a problem hiding this comment.
need to remove chat_messages
| llm: | ||
| model: "/data/models/Qwen/Qwen3.6-27B" | ||
| tensor_parallel_size: 2 | ||
| tensor_parallel_size: 4 |
There was a problem hiding this comment.
the default configure file in models folder is the default settings we got from dev.
| @@ -38,6 +38,18 @@ serve: | |||
| max_tokens: 256 | |||
| chat_messages: | |||
There was a problem hiding this comment.
need to remove chat_messages as we have chat_cases
| max_tokens: 256 | ||
| chat_messages: | ||
| - role: "user" | ||
| content: "Where is the capital of France" |
There was a problem hiding this comment.
need to remove chat_messages
| # Use this for parameters that differ from the base model YAML. | ||
| device_overrides: | ||
| my_device: | ||
| qwen_case_name: |
There was a problem hiding this comment.
explain the reason why we add this test case
There was a problem hiding this comment.
explain the reason why we add this test case
There was a problem hiding this comment.
explain the reason why we add this test case
| case, | ||
| platform, | ||
| device, | ||
| platforms_dir, |
There was a problem hiding this comment.
do we really need this parameter
|
filed an issue to track the ascend rotary operator/runtime problem: #356 based on the log, this issue is triggered when the ascend functional test uses the vendor implementation of rotary_embedding and hits an ATB RopeOperation setup failure. This is unrelated to the qwen3_6 config unification itself, so I will keep this pr focused on the qwen3_6 config changes and avoid introducing ascend operator-level changes |
| e2e: | ||
| serving: | ||
| qwen3_6: ["27b_tp2_262k", "35b_a3b_tp2_262k"] | ||
| qwen3_6: ["27b_tp2_eager", "35b_a3b_tp2_eager"] |
| - name: Check device availability | ||
| run: bash .github/scripts/${{ inputs.platform }}/check.sh | ||
|
|
||
| - name: Prepare host models |
2 participants
Summary
This PR consolidates the pending MUSA S5000, MetaX graph serving, and Ascend 910C CI updates, and keeps Qwen3.6 test configs unified across platforms.
Changes
tests/platforms/*yamldevice_overrides.Notes
This PR consolidates the relevant changes from #314, #327, and #328. The old platform-specific Qwen3.6 YAML files are not kept; their differences are represented through unified cases plus platform
device_overrides.Validation