Tsingmicro add txda backend to vllm-fl plugin. by tsingmicro-public-e · Pull Request #52 · flagos-ai/vllm-plugin-FL

tsingmicro-public-e · 2026-02-16T13:16:58Z

PR Category

PR Types

PR Description

CLAassistant · 2026-02-16T13:17:06Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ tsingmicro-public-e
❌ ruanchunyang

ruanchunyang seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

Copilot

Pull request overview

This PR introduces an initial “txda” vendor backend into the vllm-plugin-FL dispatch system and enables static-graph mode for the txda device type.

Changes:

Extend static-graph support gating to include txda.
Add a new TxdaBackend backend class with operator entrypoints (silu_and_mul, rms_norm, rotary_embedding, attention_backend).
Add the txda vendor package initializer.

Reviewed changes

Copilot reviewed 1 out of 3 changed files in this pull request and generated no comments.

File	Description
`vllm_fl/platform.py`	Adds `txda` to the static-graph enablement allowlist.
`vllm_fl/dispatch/backends/vendor/txda/txda.py`	Introduces the Txda backend class and operator/attention backend routing.
`vllm_fl/dispatch/backends/vendor/txda/__init__.py`	Exposes the new backend from the `txda` vendor package.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

tengqm · 2026-02-21T10:57:04Z

vllm_fl/dispatch/backends/vendor/txda/register_ops.py

+        #     fn=_bind_is_available(backend.rotary_embedding, is_avail),
+        #     vendor="txda",
+        #     priority=BackendPriority.VENDOR,
+        # ),


Why are we dropping the other implementations?

tengqm

Please label this PR with [WIP] if it is still a work in progress,
in case it gets merged accidentally.

tengqm · 2026-03-07T00:30:26Z

vllm_fl/dispatch/backends/vendor/txda/__init__.py

+
+from .txda import TxdaBackend
+
+__all__ = ["TxadBackend"]


TxdaBackend typo

tengqm · 2026-03-07T00:30:54Z

vllm_fl/dispatch/backends/vendor/txda/register_ops.py

+# Copyright (c) 2026 BAAI. All rights reserved.
+
+"""
+METAX backend operator registrations.


Suggested change

METAX backend operator registrations.

TsingMicro backend operator registrations.

tengqm · 2026-03-07T00:31:08Z

vllm_fl/dispatch/backends/vendor/txda/register_ops.py

+"""
+METAX backend operator registrations.
+
+This module registers all VENDOR (METAX) implementations.


Suggested change

This module registers all VENDOR (METAX) implementations.

This module registers all VENDOR (TsingMicro) implementations.

tengqm · 2026-03-07T00:31:57Z

vllm_fl/dispatch/backends/vendor/txda/register_ops.py

+
+def register_builtins(registry) -> None:
+    """
+    Register all METAX (VENDOR) operator implementations.


Suggested change

Register all METAX (VENDOR) operator implementations.

Register all TsingMicro (VENDOR) operator implementations.

tengqm · 2026-03-07T00:32:59Z

vllm_fl/dispatch/backends/vendor/txda/txda.py

+
+import torch
+
+# from vllm_fl.dispatch.backends.flaggems import FlagGemsBackend


Remove this line.

tengqm · 2026-03-07T00:33:57Z

vllm_fl/dispatch/backends/vendor/txda/txda.py

+                # Check for torch_npu (Txda PyTorch extension)
+                import torch_txda
+
+                # Check if NPU device is available


Suggested change

# Check if NPU device is available

# Check if TsingMicro txda device is available

tengqm · 2026-03-07T00:37:59Z

vllm_fl/dispatch/backends/vendor/txda/txda.py

+                if torch.txda.is_available() and torch.txda.device_count() > 0:
+                    TxdaBackend._available = True
+                else:
+                    TxdaBackend._available = False


It is weird that we change class properties in an instance method.
You may want to annotate this method with @classmethod.

tengqm · 2026-03-07T00:40:29Z

vllm_fl/distributed/device_communicators/flagcx.py

                flagcx_path = os.getenv('FLAGCX_PATH')
-                library_path=os.path.join(flagcx_path, "build/lib/libflagcx.so")
+                #library_path=os.path.join(flagcx_path, "libflagcx.so") # rcy fix
+                library_path= "/usr/local/kuiper/lib/libflagcx.so"


We are using hard-coded path rather than using the environment variable?

tengqm · 2026-03-07T00:42:25Z

vllm_fl/ops/fla/fused_recurrent.py

 from vllm_fl.utils import use_flaggems_op

-if use_flaggems_op("fused_recurrent_gated_delta_rule_fwd"):
+if True:


Why this change?

tengqm · 2026-03-07T00:44:33Z

vllm_fl/platform.py

-    ray_device_key: str = "GPU"
+    ray_device_key: str = "flagos"
    dist_backend: str = "flagcx" if "FLAGCX_PATH" in os.environ else "nccl"
+    device_control_env_var: str = "TXDA_VISIBLE_DEVICES"


This is not a platform specific class ...
This property should not live here?

vllm_fl/dispatch/backends/vendor/txda/__init__.py

+
+from .txda import TxdaBackend
+
+__all__ = ["TxadBackend"]


vllm_fl/dispatch/backends/vendor/txda/txda.py

+
+from __future__ import annotations
+
+from typing import Optional, Union


vllm_fl/distributed/device_communicators/flagcx.py

                                flagcx_stream)
        self.flagcx.adaptor_stream_free(flagcx_stream)
+        if change_type:
+            in_tensor = in_tensor.to(torch.bfloat16)


vllm_fl/models/qwen3_5.py

vllm_fl/ops/fla/fused_recurrent.py

        fused_recurrent_gated_delta_rule_fwd,
    )
-
+from flag_gems.fused.FLA.utils import input_guard


vllm_fl/worker/worker.py

+        # if self.init_snapshot.free_memory < self.requested_memory:
+        #     GiB = lambda b: round(b / GiB_bytes, 2)
+        #     raise ValueError(
+        #         f"Free memory on device "
+        #         f"({GiB(self.init_snapshot.free_memory)}/"
+        #         f"{GiB(self.init_snapshot.total_memory)} GiB) on startup "
+        #         f"is less than desired GPU memory utilization "
+        #         f"({self.cache_config.gpu_memory_utilization}, "
+        #         f"{GiB(self.requested_memory)} GiB). Decrease GPU memory "
+        #         f"utilization or reduce GPU memory used by other processes."


cyber-pioneer · 2026-03-09T03:09:39Z

vllm_fl/platform.py

    torch_device_fn = device_info.torch_device_fn
    vendor_name = device_info.vendor_name
-    ray_device_key: str = "GPU"
+    ray_device_key: str = "flagos"


why modify here?

cyber-pioneer · 2026-03-09T03:09:55Z

vllm_fl/platform.py


    def is_cuda_alike(self) -> bool:
        """Stateless version of [torch.cuda.is_available][]."""
-        if self.vendor_name == "iluvatar":


why remove here?

Tsingmicro add txda backend to vllm-fl plugin.

96a23a1

github-actions bot added the core label Feb 16, 2026

aoyulong requested a review from Copilot February 20, 2026 00:38

Copilot started reviewing on behalf of aoyulong February 20, 2026 00:38 View session

Copilot AI reviewed Feb 20, 2026

View reviewed changes

tsingmicro-public-e and others added 2 commits February 20, 2026 10:55

Merge branch 'flagos-ai:main' into main

befc101

add tsingmicro txda backend code to vllm-fl.

c716683

tengqm reviewed Feb 21, 2026

View reviewed changes

tsingmicro-public-e and others added 3 commits February 26, 2026 14:57

Merge branch 'flagos-ai:main' into main

56e2f3f

Merge branch 'flagos-ai:main' into main

6cec619

support qwen3.5 module in vllm-fl for txda.

ac17633

tengqm reviewed Mar 7, 2026

View reviewed changes

github-advanced-security bot found potential problems Mar 7, 2026

View reviewed changes

cyber-pioneer reviewed Mar 9, 2026

View reviewed changes

	METAX backend operator registrations.
	TsingMicro backend operator registrations.

	This module registers all VENDOR (METAX) implementations.
	This module registers all VENDOR (TsingMicro) implementations.

	Register all METAX (VENDOR) operator implementations.
	Register all TsingMicro (VENDOR) operator implementations.


		import torch

		# from vllm_fl.dispatch.backends.flaggems import FlagGemsBackend

	# Check if NPU device is available
	# Check if TsingMicro txda device is available


		from __future__ import annotations

		from typing import Optional, Union

Conversation

tsingmicro-public-e commented Feb 16, 2026

PR Category

PR Types

PR Description

Uh oh!

CLAassistant commented Feb 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tengqm left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Check failure

Check notice

Check notice

Check notice

Check notice

Check notice

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

CLAassistant commented Feb 16, 2026 •

edited

Loading