RocmPlatform::needs_weight_staging() is stale-false, so the #1123/#1870 device-fit refusal never runs on ROCm

Row: BACKEND-ROCM

RocmPlatform::needs_weight_staging() (rocm.cpp:88)
hardcodes false, with a comment dated to the W0 skeleton: "in W0 there is one
registered op, so the only path that can run at all is the host-resident one
... a discrete AMD card will eventually answer true ... Revisit at M2."

That comment is stale. Since W0, ROCm has landed grouped/non-grouped keep-quant
expert GEMM (#523), MoE combine/gate ops (#509), decode-skinny wvSplitK GEMMs
(#506), self-registered ROCM_ATTN, and hipGraph decode capture (W1). Measured
on this box (gfx1200), tests/vt/test_backend_cross_device runs real ROCm
MoE kernels against the CPU oracle within tolerance today. The platform is not
"one registered op" anymore.

Why this is more than a stale comment

needs_weight_staging() gates the ONE production call site of
CheckDeviceWeightFit (model_loader.cpp:2273),
the load-time refusal ENG-EXPERT-STREAM (#1123) added specifically so a
checkpoint that cannot fit device memory is refused BY NAME before any
allocation, instead of dying mid-load with a raw cudaMalloc/hipMalloc: out of memory. CheckDeviceWeightFit returns immediately when
needs_weight_staging() is false — nothing is even computed.

Measured directly: on this box, loading a checkpoint through vllm-cli --device auto with VT_DEVICE_WEIGHT_BUDGET_BYTES=1 (a budget nothing can
satisfy) produced no refusal at all and proceeded to the next load stage.
Meanwhile the actual device weight allocation this refusal exists to guard —
d.b.Alloc(nb) in qwen3_5.cpp's ResidentWeight construction — is not
gated by needs_weight_staging(); it allocates unconditionally on whatever
device the resolved queue targets. So on ROCm today: the crash #1870 reports
(vt rocm: hipMalloc: out of memory) is real and reachable, and the refusal
meant to replace it with a named message is not, because the guard that would
run it believes this platform never stages weights.

Relationship to #1870

#1870 reports the crash and
names the fix as "resolve the expanded residency requirement at load, compare
it against the device budget, and refuse." That arithmetic fix landed
separately (policy_forces_full_expand in gguf_device_fit.h/.cpp,
.agents/specs/gguf-device-fit-expand-policy.md) and is necessary — the
pre-existing min()-based bound under-counted a forced-expand load by
~4x — but it is not sufficient: with needs_weight_staging() false, the
corrected bound is still never evaluated on ROCm. #1870 stays open until this
row closes the gap.

Scope

Not proposed as a design, only to bound it: flipping needs_weight_staging()
(and giving residency_policy() a real, probed device_memory_total_bytes
instead of the current default {}) is a platform-capability graduation, not
a one-line flag flip. It is also the ONE input to at least these other
decisions, all of which currently take the "not staging" branch on ROCm and
would move:

• DirectDeviceLoadEligible (qwen3_5_dense_weights.cpp:125) — load-path optimization (straight-to-device vs. via host)
• IndexedGdnStateIoEnabled (qwen3_5.cpp:3354) — GDN state-cache kernel dispatch; ROCm currently gets the CPU-style row-copy reference path by default because this flag is false
• MergedGdnBaEnabled / MergedGdnQkvzEnabled's eligibility checks (qwen3_5.cpp:3573, 3881) — merged-projection GDN kernel eligibility
• PackedGdnDecodeEligibility (qwen3_5.cpp:3910) and PrepareGdnFp8Resident/PrepareBf16Resident (qwen3_5.cpp:8478, 8498) — decode-path prep gating

Each of those needs its own correctness check once the flag moves — several
change ROCm's default kernel dispatch, not just an opt-in path, so a token-exact
verification against the pinned oracle is needed per affected model, not just a
build-and-run.

History

Found while implementing #1870's arithmetic fix and trying to reproduce the
crash end-to-end on real gfx1200 hardware: the corrected bound never fired
because the guard around it short-circuits first. Filed separately per
AGENTS.md's owed-row rule — this is a platform-capability change with its own
review burden, not a slice of the arithmetic fix.

Owned by BACKEND-ROCM.

Start with src/vllm/platforms/rocm.cpp:88 and the call at src/vllm/entrypoints/model_loader.cpp:2273, then trace residency_policy() and CheckDeviceWeightFit. Review the listed qwen3_5 and qwen3_5_dense_weights.cpp eligibility sites, plus gguf_device_fit and tests/vt/test_backend_cross_device. Done means ROCm has a probed memory budget, the refusal runs, and affected model paths pass token-exact oracle checks.