From 683349864680f07f72c58a1aaec7c88f48a76110 Mon Sep 17 00:00:00 2001 From: Xinyuan Tong Date: Tue, 22 Sep 2026 07:53:54 +0000 Subject: [PATCH] model: prune comments and redundant tests in dsv41 vision CP --- python/sglang/srt/models/deepseek_v4.py | 2 -- .../unit/managers/test_mm_owner_embedding.py | 8 +------- .../models/test_deepseek_v41_vision_cp_inputs.py | 15 --------------- .../unit/server_args/test_server_args.py | 8 -------- 4 files changed, 1 insertion(+), 32 deletions(-) diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index 35f3e9ae7..d20efaeb0 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -4922,8 +4922,6 @@ class DeepseekV4ForCausalLM(nn.Module): self.image_start = nn.Parameter(torch.empty(config.hidden_size)) self.image_end = nn.Parameter(torch.empty(config.hidden_size)) self.image_newline = nn.Parameter(torch.empty(config.hidden_size)) - # Ranks of this group run identical image chunks; one owner encodes each - # span and broadcasts it. None keeps the replicated encoder. self.mm_owner_group = ( select_owner_group(get_parallel()) if self.vision is not None and _is_cuda diff --git a/test/registered/unit/managers/test_mm_owner_embedding.py b/test/registered/unit/managers/test_mm_owner_embedding.py index e4ef8b00c..5133bc68f 100644 --- a/test/registered/unit/managers/test_mm_owner_embedding.py +++ b/test/registered/unit/managers/test_mm_owner_embedding.py @@ -229,8 +229,6 @@ class _VisionStub(DeepseekV4ForCausalLM): class _TracedGroup: - """Records every collective (op, source, shape, group size) around a real coordinator.""" - def __init__(self, inner): self.inner = inner self.trace = [] @@ -441,7 +439,6 @@ def _topology_program(rank, world_size): singles = _coordinator([[r] for r in range(8)], rank) out = {} - # TP8, DP1, CP off. tp_group, attn_tp, attn_cp = ( _TracedGroup(tp8), _TracedGroup(tp8), @@ -469,7 +466,6 @@ def _topology_program(rank, world_size): assert torch.equal(embeds, _expected_embeds(embed, requests, {x: 0, y: 5, z: 2})) out["cp1"] = {"encoded": list(model.encoded), "trace": list(attn_tp.trace)} - # TP8, DP1, CP8 through the CP runner. tp_group, attn_tp, attn_cp = ( _TracedGroup(tp8), _TracedGroup(singles), @@ -516,7 +512,7 @@ def _topology_program(rank, world_size): finally: init_cp_strategy(enable_prefill_cp=False, cp_size=1, cp_strategy="interleave") - # TP8, attention-DP2, CP off: one replica is text-only first. + # One attention-DP replica is text-only first. tp_group, attn_tp, attn_cp = ( _TracedGroup(tp8), _TracedGroup(replicas), @@ -804,7 +800,6 @@ def test_owner_actions_and_cache_lifetime_across_asymmetric_ranks(): def test_tp8_cp1_and_cp8_dedupe_and_attention_dp_replicas_stay_isolated(): - """One encode per key on TP8 with CP off and CP8; attention-DP replicas never share a collective.""" results = _run_ranks(8, _topology_program) cp1 = [r["cp1"] for r in results] @@ -839,7 +834,6 @@ def test_tp8_cp1_and_cp8_dedupe_and_attention_dp_replicas_stay_isolated(): def test_failures_agree_before_payload_text_embedding_or_body(): - """Every rank raises the same error and none enters a later collective.""" results = _run_ranks(4, _failure_program) cases = ( "prepare", diff --git a/test/registered/unit/models/test_deepseek_v41_vision_cp_inputs.py b/test/registered/unit/models/test_deepseek_v41_vision_cp_inputs.py index 748d3cc22..35082a7c3 100644 --- a/test/registered/unit/models/test_deepseek_v41_vision_cp_inputs.py +++ b/test/registered/unit/models/test_deepseek_v41_vision_cp_inputs.py @@ -200,21 +200,6 @@ class TestDeepseekV41VisionPrefillCPInputs(CustomTestCase): input_embeds=input_embeds, ) - def test_prepare_model_inputs_merges_on_full_layout(self): - forward_batch, item = _build_batch() - scheduler_ids = forward_batch.input_ids.clone() - - model_ids, embeds = self._prepare(forward_batch) - - self.assertTrue(torch.equal(forward_batch.input_ids, scheduler_ids)) - self.assertIs(forward_batch.mm_input_embeds, embeds) - self.assertTrue(torch.equal(model_ids, _canonical(scheduler_ids))) - self.assertTrue(torch.equal(embeds[IMAGE_ROWS], _image_span(item)[1:7])) - text_rows = model_ids != IMAGE_TOKEN_ID - with torch.no_grad(): - text_embeds = self.embed(scheduler_ids[text_rows]) - self.assertTrue(torch.equal(embeds[text_rows], text_embeds)) - def test_cp_runner_merges_before_shard(self): runner = EagerRunner.__new__(EagerRunner) runner.model_runner = SimpleNamespace(model=self.model) diff --git a/test/registered/unit/server_args/test_server_args.py b/test/registered/unit/server_args/test_server_args.py index 90dd7be50..36e8880a7 100644 --- a/test/registered/unit/server_args/test_server_args.py +++ b/test/registered/unit/server_args/test_server_args.py @@ -4123,14 +4123,6 @@ class TestDeepseekV41VisionPrefillCPArgs(CustomTestCase): ): handle_model_specific_adjustments(args) - def test_interleave_eager_prefill_is_accepted(self): - args = self._args() - validate_deepseek_v41_features(args) - self.assertEqual( - resolution_result(args, "cuda_graph_config").prefill.backend, - Backend.DISABLED, - ) - def test_zigzag_is_rejected_only_with_vision(self): with self.assertRaisesRegex(ValueError, "requires --cp-strategy interleave"): validate_deepseek_v41_features(self._args(cp_strategy="zigzag"))