HIP graph replay reads stale device memory for externally-written buffer on gfx950 (MI355X) — causes MEMORY_VIOLATION in LLM decode

Environment

• GPU: AMD MI355X (gfx950)
• ROCm: 7.2.x (rocm7.2.4)
• PyTorch: 2.10.0+rocm7.2.4
• Setup: 8-GPU tensor-parallel LLM inference (TP=8) using PyTorch torch.cuda.CUDAGraph for decode

---

Bug Summary

When a device buffer is written via H2D copy on the default stream immediately before graph.replay() — also on the same default stream — the in-graph kernel reads a stale value from the previous step instead of the freshly written value.

This is not a stream-ordering bug (the H2D copy and replay are on the same stream in program order), not a dangling-pointer bug (the captured pointer equals the live buffer address), and not a concurrent-write bug (the buffer is arange both before and after replay). The stale read is localized to the HIP graph replay path itself.

---

Symptom

Intermittent Memory access fault by GPU node-N MEMORY_VIOLATION during LLM decode (CUDAGraph replay path), non-deterministic, absent in eager mode. Triggered by a small integer buffer (cu_seqlens_q, int32, ~513 elements) that is written correctly each step but read stale inside the graph.

Debug-agent wave dump at fault:

wave_NNNN: pc=0x... <_convert_req_index_to_global_index_kernel> (stopped, reason: MEMORY_VIOLATION)
scalar registers:
    s3: 000003e3    ← qo_end = 0x3e3 = 995  (STALE cumulative value from prior prefill step)
    s2: 0000008b    ← token_id = 139  (loop variable, should be ≤ 256 for decode)
    ...
vector registers:
    v3: [0] 0000772b [1] 0000772b ...   ← derived offset ≈ 29 000+, far OOB in token_indices

The buffer should contain arange(0, 257) (decode: 1 token per request). Instead the graph kernel read qo_end = 995, the cumulative value left by the immediately preceding prefill step.

Source location: for token_id in range(qo_start, qo_end) — the stale qo_end drives token_id to ~995, causing an out-of-bounds load in token_indices[token_id].

---

Evidence: buffer is physically correct; stale is in the graph's read path

We performed three independent probes to localize the stale:

1. Before-replay D2D snapshot (no host sync)

Right before graph.replay(), a GPU→GPU copy (non_blocking=True, same default stream) snapshots cu_seqlens_q into a persistent buffer. A daemon thread later saves it — no synchronization on the hot path.

Result (8 ranks × 3000 steps, warmup + steady-state): buffer is always arange(0, 257), max = 256, zero cumulative rows. The physical device memory is correct before replay.

2. After-replay D2D snapshot (rolling ring, no host sync)

Same approach, taken after replay, rolling ring of 512 steps overwritten every 100 steps.

Result (8 ranks × 512 steps covering the crash step): buffer is always arange(0, 257), max = 256, zero cumulative rows — even at the exact steps preceding the crash.

3. In-graph kernel register dump (ROCm debug agent)

HSA_TOOLS_LIB=librocm-debug-agent.so.2, graph mode preserved (not eager), --save-code-objects. Crash reproduced with the unmodified kernel; wave dump captured at fault.

Result: s3 = 0x3e3 = 995 = stale qo_end from the immediately preceding prefill step (which legitimately wrote cumulative ~988 into the buffer before the current decode step overwrote it with arange). The buffer is provably arange (probes 1 + 2), yet the graph kernel read 995.

Conclusion from probes

Probe	Buffer content	Kernel read
Before-replay D2D	arange (max 256)	—
After-replay D2D	arange (max 256)	—
Debug-agent register	—	995 (stale cumulative)

Same address (ADDR_PROBE confirmed capture pointer == live pointer at every replay). Same physical memory (D2D probes confirm arange both sides of replay). Different value inside the graph. The stale is 100% in the graph replay's memory read path.

---

Characteristics

• CUDAGraph required: eager mode runs 145 000+ decode steps with zero violations.
• Multi-GPU TP=8 required: we attempted single-GPU minimal repros (torch copy graph, triton kernel graph, 500-kernel large graph, shared-pool multi-graph, all with external H2D write before replay) — none reproduced. The bug requires the real TP=8 setup with RCCL collectives inside the graph.
• Non-deterministic / flaky: same commit alternates pass/fail; concurrent-scale-dependent (higher batch = more prefill-decode interleaving = higher stale probability).
• One-epoch lag: the stale value is specifically the value written by the immediately preceding prefill step — not the capture-time value, not an arbitrary old value. This suggests the graph's view of the buffer lags by one write epoch.
• In-kernel instrumentation masks the bug (Heisenbug): adding a tl.store to record the loaded value (and clamping the loop to avoid the crash) causes the bug to disappear — the recompiled/recaptured kernel stops exhibiting the stale. This makes in-kernel probes unreliable for this bug; the debug-agent register dump (which does not modify the kernel binary) is the only reliable in-kernel observable.
• Graph pool: multiple graphs (one per decode batch size) share a single graph_pool_handle.

---

H2D write details

The write path:

# CpuGpuBuffer.copy_to_gpu()
self.gpu.copy_(self.cpu, non_blocking=True)   # pinned H2D on current stream

Called in prepare_inputs on the default stream, immediately before graph.replay() also on the default stream. Program-order synchronization should guarantee visibility. The forward pass is single-stream synchronous with no side-stream switching.

---

Ask

1. Is this a known HIP-graph limitation on gfx950 (e.g., graph replay does not guarantee observability of H2D writes on the same stream made between replays)?
2. Is there a timeline for a fix in HIP runtime or the graph replay path?
3. Is there a recommended pattern to force the graph to observe external writes (other than inserting a cudaStreamSynchronize between the write and replay, which would defeat the performance purpose of CUDAGraph)?

No repository file or test is named. Start with the H2D path in CpuGpuBuffer.copy_to_gpu(), its prepare_inputs caller, and graph.replay() on the default stream; then reproduce the TP=8 gfx950 case while tracing HIP graph replay visibility. Done means the limitation or regression is localized and a supported fix or synchronization pattern is identified.