CP modeling: unify device_per_node and bw_intra/bw_inter data sources (follow-up PR)

Follow-up from R29's review on the CP modeling PR. The three items below are deferred to separate fix/bug/feature PRs — they require deeper refactoring of the Device type and level_bandwidth semantics, which is out of scope for the current CP modeling review. Filing this issue to track them and address them together once the Device/level_bandwidth structure is fully understood.

1. Unify device_per_node source

device_per_node in cp_comm_layer_detailed and related paths should be derived from machine.device.intra_node_num() instead of the flat field on ccfg.

2. Recalibrate bw_intra / bw_inter dimensionality

ccfg.bw_intra = 400 and Device_A2.level_bandwidth[0] = 50 are not the same quantity:

• level_bandwidth = [50, 10] is per-link physical bandwidth (introduced by ch-l in 0f0e617a1 / 738c37f5a, the Type abstraction treats it as per-link).
• ccfg.bw_intra = 400 is the effective/aggregate bandwidth after 8-card HCCS aggregation (calibrated value).

Directly substituting one for the other would inflate intra-node CP communication time by ~8x and distort the search-space ranking. The level_bandwidth semantics (per-link vs aggregate) must be recalibrated before deciding whether to replace ccfg.bw_intra/bw_inter with values from level_bandwidth.

3. Consolidate the three bandwidth sources into a single source of truth

Currently there are three independent bandwidth sources:

• Production path: cp_comm_layer_detailed → _cp_resolve_topology reads ccfg.bw_intra / ccfg.bw_inter (flat fields, defaults 400 / 25).
• Test path: get_cp_topology hardcodes 300 / 25.
• hardware.py: Device_A2.level_bandwidth = [50, 10].

Unifying these three into a single source of truth has impact beyond the CP modeling review scope, so it will be handled as a separate PR.

Dependency: Item (1) is a prerequisite for (2) and (3) — the device_per_node source must be unified first before the bandwidth dimensionality can be recalibrated and the sources consolidated. The three items will be addressed together in one follow-up PR.

schema_version: 1
source: gitcode
gitcode_repo: mindspore/hyper-parallel
gitcode_issue: 306
source_url: https://gitcode.com/mindspore/hyper-parallel/issues/306

Start by tracing device_per_node through cp_comm_layer_detailed and _cp_resolve_topology, then compare get_cp_topology with hardware.py's Device_A2.level_bandwidth. Read the Device, intra_node_num(), and level_bandwidth semantics before changing the bandwidth sources. Done means device_per_node is unified first, bandwidth dimensionality is recalibrated, and production, test, and hardware paths share one source of truth.