ci: derive the GPU test legs from pyproject; add dev-latest

[aviv1ron1]

The GPU workflow no longer hardcodes its vLLM matrix. A first job reads
this commit's pyproject.toml and runs up to two legs, from the groups
dev and dev-latest, each labelled after the vLLM minor it pins (one
leg if both pin the same minor). A vLLM bump is now a pyproject edit:
add the new vllmNN / dev-vllmNN groups and repoint dev-latest.

dev-latest is an alias of dev-vllm27, with the same conflicts declared.
uv.lock is relocked with the repo's pinned uv 0.8.4; no package version
changes, the diff is uv's resolution markers now naming dev-latest.

Also points /gpu-test-dev at tests/vllm/test_multi_switch.py:
test_single_switch.py was removed, so that suite collected nothing.

Signed-off-by: aviv ron rona@il.ibm.com

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[aviv1ron1]

/gpu-test-dev

[github-actions]

✅ GPU tests passed — vllm26-dev · ♻️ cached

42 passed in 32.61s

♻️ Reused result — the tests did not run again.
This commit was already tested with this exact scope, and the result above
is that run's, finished 2026-10-10T15:30:55+00:00.
The run that produced it: https://github.com/generative-computing/granite-switch/actions/runs/38063754729

Re-comment the command with --no-cache to test this commit again.

Commit: bf3655ad3e4a237f9b0e43248130d4eebcf40065
Full run & artifact log

Last 40 log lines

tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_non_control_all_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_unregistered_tokens_produce_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_control_token_at_position_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_one_text PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_one_control PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_two_control_text PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_two_text_control PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_long_all_text_produces_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_duplicate_same_adapter PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[1] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[2] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[5] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[10] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[20] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[50] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[100] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[3] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[8] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[33] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[64] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[129] PASSED
tests/vllm/test_multi_switch.py::TestGeometry::test_num_cache_and_attn_layers[multi] PASSED
tests/vllm/test_multi_switch.py::TestGeometry::test_coded_builds_two_attention_layers PASSED
tests/vllm/test_multi_switch.py::TestBatchlessSingleRequest::test_base_a_b_flat[multi] PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_matches_individual_no_base PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_return_to_base_isolated PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_second_request_anchor_independent PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_single_request_batch_equals_batchless PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_many_controls_vs_ground_truth PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_ground_truth_survives_request_order PASSED
tests/vllm/test_multi_switch.py::TestCountingCeiling::test_bound_is_safe_on_the_real_engine PASSED
tests/vllm/test_multi_switch.py::TestCountingCeiling::test_aliasing_does_appear_above_the_bound 
  guard admits          188
  first wrong at        204 control tokens
  first wrong position  405 (got 2, wanted 1)
PASSED

============================= 42 passed in 32.61s ==============================

Verdict: PASSED

[github-actions]

✅ GPU tests passed — vllm27-dev · ♻️ cached

42 passed in 31.73s

♻️ Reused result — the tests did not run again.
This commit was already tested with this exact scope, and the result above
is that run's, finished 2026-10-10T15:31:07+00:00.
The run that produced it: https://github.com/generative-computing/granite-switch/actions/runs/38063754729

Re-comment the command with --no-cache to test this commit again.

Commit: bf3655ad3e4a237f9b0e43248130d4eebcf40065
Full run & artifact log

Last 40 log lines

tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_non_control_all_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_unregistered_tokens_produce_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_control_token_at_position_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_one_text PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_one_control PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_two_control_text PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_seq_len_two_text_control PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_long_all_text_produces_zero PASSED
tests/vllm/test_multi_switch.py::TestEdgeCasesCoded::test_duplicate_same_adapter PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[1] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[2] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[5] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[10] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[20] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[50] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_matches_input[100] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[3] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[8] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[33] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[64] PASSED
tests/vllm/test_multi_switch.py::TestShapeCoded::test_output_length_with_transitions[129] PASSED
tests/vllm/test_multi_switch.py::TestGeometry::test_num_cache_and_attn_layers[multi] PASSED
tests/vllm/test_multi_switch.py::TestGeometry::test_coded_builds_two_attention_layers PASSED
tests/vllm/test_multi_switch.py::TestBatchlessSingleRequest::test_base_a_b_flat[multi] PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_matches_individual_no_base PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_return_to_base_isolated PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_second_request_anchor_independent PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_single_request_batch_equals_batchless PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_many_controls_vs_ground_truth PASSED
tests/vllm/test_multi_switch.py::TestContinuousBatching::test_batch_ground_truth_survives_request_order PASSED
tests/vllm/test_multi_switch.py::TestCountingCeiling::test_bound_is_safe_on_the_real_engine PASSED
tests/vllm/test_multi_switch.py::TestCountingCeiling::test_aliasing_does_appear_above_the_bound 
  guard admits          188
  first wrong at        204 control tokens
  first wrong position  405 (got 2, wanted 1)
PASSED

============================= 42 passed in 31.73s ==============================

Verdict: PASSED