Revert "A balanced traffic pattern for AG minimal. (#36607)"

[rdraskicTT]

Causes incorrect outputs and hangs for llama models
Example failing runs https://github.com/tenstorrent/tt-metal/actions/runs/21972598335/job/63477567742

• llama 8b galaxy demo
• llama 8b vllm nightly

[Copilot]

Pull request overview

Reverts the previously introduced “balanced traffic pattern” behavior in the minimal all-gather-async implementation, motivated by incorrect outputs and hangs observed in Llama-model workloads.

Changes:

• Removes split-forwarding (“balanced traffic”) logic from minimal all-gather async reader/writer device kernels.
• Removes a fabric packet-size vs page-size TT_FATAL guard in the minimal default program factory.
• Deletes a WAN-focused ring perf/trace test block and related helper/imports from the nightly test.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
ttnn/cpp/ttnn/operations/experimental/ccl/all_gather_async/device/kernels/minimal_default_writer.cpp	Reverts split-forwarding logic and restores simpler per-slice forwarding behavior.
ttnn/cpp/ttnn/operations/experimental/ccl/all_gather_async/device/kernels/minimal_default_reader.cpp	Reverts split-forwarding logic and simplifies forwarding/wait behavior accordingly.
ttnn/cpp/ttnn/operations/experimental/ccl/all_gather_async/device/all_gather_async_default_program_factory.cpp	Removes validation tying fabric packet payload size to tensor page size.
tests/nightly/tg/ccl/test_minimal_all_gather_async.py	Removes WAN ring perf/trace test and related helper/imports; leaves some imports now unused.

Comments suppressed due to low confidence (1)

ttnn/cpp/ttnn/operations/experimental/ccl/all_gather_async/device/all_gather_async_default_program_factory.cpp:393

• Removing the packet_size_bytes >= page_size validation can lead to num_pages_per_packet == 0, which makes num_tiles_to_write_per_packet == 0. That will result in zero-sized CBs and/or infinite loops in the reader/writer kernels (tiles_read never advances), i.e. hangs. Please restore the guard (TT_FATAL) or otherwise enforce num_tiles_to_write_per_packet >= 1 with a clear error when the fabric packet payload is smaller than a tensor page size.

    // L1 Scratch CB Creation
    const size_t packet_size_bytes = tt::tt_fabric::get_tt_fabric_channel_buffer_size_bytes();
    uint32_t l1_scratch_cb_page_size_bytes = page_size;

    // scatter-write currently supports 4 distinct noc addresses
    uint32_t max_target_noc_addresses_per_packet = 4;

    // for bfloat8_b, tile_num_per_link=6, we would need to send 2 packages, but they can be of size 3 instead of 4
    uint32_t num_pages_per_packet = packet_size_bytes / l1_scratch_cb_page_size_bytes;
    uint32_t num_tiles_to_write_per_packet = std::min(max_target_noc_addresses_per_packet, num_pages_per_packet);
    uint32_t cb_num_pages = 3 * num_tiles_to_write_per_packet;  // triple buffering

[Copilot]

comp_equal and comp_pcc are imported but only referenced in commented-out code, so they are currently unused. If the repo runs ruff/flake8 in CI, this will fail with an unused-import error; please remove the unused imports or re-enable the assertions that use them.

Suggested change

from tests.tt_eager.python_api_testing.sweep_tests.comparison_funcs import comp_equal, comp_pcc

[rdraskicTT]

@llongTT fyi

[dpopovTT]

/codeowners bypass

[tenstorrent-github-bot]

✅ CodeOwners bypass approval granted by @dpopovTT (metalium-developers-infra team)

[tenstorrent-github-bot]

✅ CodeOwners Bypass Approval Granted\n\nThis PR has been approved by @dpopovTT (metalium-developers-infra team) using the bypass mechanism.\n\n⚠️ Note: This bypass should only be used for emergency fixes or when standard approval process is blocked.