feat(vidur): sweep P/D network configurations

[catyans]

Summary

Add pd_networks as an optional Vidur config-explorer dimension so the existing capacity search and Pareto analyzer can compare colocated and P/D-separated deployment points across fabric bandwidths.

• introduces a validated PDNetworkConfig;
• sweeps P-node ratio, P/D bandwidth, RDMA bandwidth, NVLink bandwidth and transfer dtype;
• passes every point through existing ReplicaConfig CLI fields;
• rejects invalid ratios, non-positive bandwidths and splits that produce zero P or D replicas;
• includes non-default network points in deterministic job/cache keys;
• preserves legacy YAML behavior and legacy keys when pd_networks is absent;
• documents mixed and 1P:1D examples in debug_config.yml.

Closes #296.

Validation

Three focused tests cover design-space expansion, legacy compatibility and invalid-point rejection.

Remote smoke on lingjun-101, Python 3.11, fresh branch clone (1fca827):

• test_pd_network_dimension_expands_job_configs: PASS
• test_legacy_config_keeps_single_mixed_mode_job: PASS
• test_invalid_pd_split_or_bandwidth_is_rejected: PASS

No GPU or production service was used.

[Copilot]

Pull request overview

Adds an optional pd_networks dimension to Vidur’s config-explorer so capacity search/Pareto analysis can sweep colocated vs P/D-separated deployments across multiple fabric bandwidth/dtype points while preserving legacy behavior when the dimension is absent.

Changes:

• Introduces PDNetworkConfig and threads it through JobConfig generation, validation, and CLI-arg materialization.
• Extends job/cache key generation to incorporate the selected P/D network point (when non-default).
• Adds documented example sweep points in debug_config.yml and a focused unit-test suite.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/debug_config.yml	Adds example pd_networks sweep points for mixed vs P/D-separated runs.
vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/config.py	Implements PDNetworkConfig, integrates it into JobConfig validation/keying/config materialization, and sweeps it in config generation.
vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/init.py	Exports PDNetworkConfig from the config package.
vidur-alibabacloud/tests/test_pd_network_config_sweep.py	Adds tests for design-space expansion, legacy behavior, and invalid-point rejection.

---

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

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[catyans]

Contributor metadata follow-up: rewrote both commits without content changes so author, committer, and the sole Signed-off-by trailer use the email linked to GitHub user catyans. New head: 02acd08. The review fixes remain included and the focused suite remains 4 passed.

[catyans]

recheck

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

Suppressed comments (3)

vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/config.py:240

• generate_job_configs() uses config.get("pd_networks", ...), which will return None (or []) if the YAML includes pd_networks: but leaves it empty. That makes itertools.product(...) fail with a TypeError (or produce zero jobs) instead of falling back to the intended legacy mixed-mode default.

            config["schedulers"],
            config["tp_dimensions"],
            config["pp_dimensions"],
            config["batch_sizes"],
            config.get("pd_networks", [{"name": "mixed"}]),

vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/config.py:272

• generate_unique_model_job_configs() indexes config.get("pd_networks", ...)[0]. If pd_networks is present but empty/null in YAML, this will raise IndexError (or TypeError) during cache warmup, even though the dimension is optional.

        pd_network_config = PDNetworkConfig(
            **config.get("pd_networks", [{"name": "mixed"}])[0]
        )

vidur-alibabacloud/vidur/config_optimizer/config_explorer/config/config.py:126

• PDNetworkConfig.is_valid() validates ratio and bandwidths, but does not validate pd_p2p_comm_dtype. The underlying ReplicaConfig CLI restricts this field via choices (e.g. fp8, float16, float32), so an invalid dtype would only fail later during simulation/CLI parsing instead of being rejected up front as an invalid sweep point.

This issue also appears in the following locations of the same file:

• line 236
• line 270

    def is_valid(self, num_replicas: int):
        if not 0 < self.pd_node_ratio <= 1:
            return False
        if min(
            self.pd_p2p_comm_bandwidth,