Skip to content

Sync rdna_extras with upstream opengfx1030/vllm-rdna 3e274335, and pciescope knows the v620-Experimentation box - #5

Merged
jmacpolin merged 121 commits into
rdna_extrasfrom
claude/determined-keller-joz2xj
Nov 15, 2026
Merged

jmacpolin merged 121 commits into
rdna_extrasfrom
claude/determined-keller-joz2xj

Conversation

@jmacpolin

Copy link
Copy Markdown
Owner

Overview

This syncs rdna_extras with upstream opengfx1030/vllm-rdna rdna_extras at 3e274335 (2026-11-14): 119 commits on the same vLLM v0.33 base, taken whole. One commit of our own follows, so pciescope recognises the four-V620 box of jmacpolin/v620-Experimentation.

The owner asked for all of it (2026-11-15): "Add everything, even the stuff you said doesn't impact our serving path. New models may be added later, so use it all." Then: "I want the pciescope".

Claims

Validation

  • pciescope's own suite (Python 3.13, pytest, textual and mcp in a venv): 967 passed and 1 skipped at the merge; 968 passed and 1 skipped with 976339b3 (+1 test for the new layouts).
  • Clean merge: the trial merge of 3e274335 into rdna_extras had no conflicts, and the merged tree differs from upstream's only by our thermal-abort commits.
  • Not built here: no image, and no kernel compile for gfx1030. The jmacpolin/v620-Experimentation PR that pins this branch builds it on the host. upstream ran LeapQuant on 4× V620 (gfx1030, TP4).

Details

  • Why a merge: the same as the last sync (jmacpolin/vllm-rdna#3). The fork's own commits stay as they are, and nothing pushed is rewritten.
  • pciescope on that box: the v620 repo runs it through a wrapper, tools/pciescope.sh, which holds the v620x4 lease while no engine runs (its one-engine rule).
    • The switch reader needs root.
    • Its counter writes stay impossible on this chip until the chip's serial number is added to regs.KNOWN_CHIPS. That needs a read on the host first, and is not part of this PR.

Pull Request Checklist
  • I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

  • AI assistance was used during the creation of this PR.

  • Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

  • Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

  • Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

  • Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.


Generated by Claude Code

FaisalBiyari and others added 30 commits October 11, 2026 16:56
LLGemm1_kernel launches K/8 threads, so for K <= 8192 a wave32 block has up
to 32 waves, but the second-stage reduction reads only red_smem[..][0..15].
For K > 4096 on gfx10/gfx11/gfx12 the partials of waves 16..31 were dropped
and the result silently lost part of the dot product. Fold them into the
first 16 lanes before the shuffle. wave64 (gfx9) has at most 16 waves in
this range and is unchanged (compile-time WARP_SIZE check).

Ported from FaisalBiyari/vLLM-ALU-RDNA2@a8f8c02 (csrc/rocm/skinny_gemms.cu).
gfx10 (RDNA2) is neither __HIP__GFX9__ nor __HIP__GFX1X__, so the wvSplitK
kernels are the assert(false) stub there and any call faults the queue
(HSA_STATUS_ERROR_EXCEPTION). The Python dispatcher already routes gfx10
decode to gemv_f16_rdna2; make direct callers fail with a clear error.

Tests: skip the wvSplitK tests on gfx10, and add an LLMM1 case with K=4104
(17 waves on wave32) that exercises the second-stage partials 16..31.

Adapted from the gfx1030 wvSplitK dispatcher guard in
FaisalBiyari/vLLM-ALU-RDNA2@a8f8c02 (csrc/rocm/skinny_gemms.cu); this tree
does not build the wave32 path for gfx10, so the guard is unconditional.

Co-authored-by: FaisalBiyari <11506542+FaisalBiyari@users.noreply.github.com>
w4a8_gemm_rdna2 launches config 8 (a8_lds_k32_ag) only. Stop
instantiating the other eight explore configs, and drop the
uninstantiated probe kernel and the unused instantiate macro.

Co-authored-by: Cursor Grok 4.7 <cursoragent@cursor.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…opengfx1030#106)

- tools: delete probe_force_split.py (its env var has no reader after opengfx1030#104)
- docs: drop removed debug knobs from v0.31.0-port-notes (opengfx1030#103)
- docs: tunableop README points at the single tools/rdna2_028 helper (opengfx1030#101)
- test: GDN KKT hang tests opt in via VLLM_RUN_GDN_KKT_HANG_TEST=1 (opengfx1030#106)
…lacement solver (CPU-tested, no GPU run) [skip ci]

VLLM_EXPERT_STATS_DIR turns on a cudagraph-safe per-expert routing counter at the router output (one is-not-None test per MoE call and per step when off). tools/rdna/hetero_moe converts the stats, solves a hub/V620 placement for the 2 x (1 W7800 + 4 V620) box, validates plans and simulates them. Sample plan from synthetic stats; doc docs/rdna2/expert-stats-2026-11-07.md.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…t model, box budget defaults [skip ci]

Adds island_mode partition (each expert in one island, global owner indices, per-island solve), the cost model constants and fixed terms from the design spec, a spill-aware simulated step time, --mtp-expert-bytes with MTP policy hub by default, W7800 48 GB and V620 26 GiB budget defaults, and a partition sample plan.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…P rank (rank agreement check) [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…llector on V620 slot B [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…ip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…ecks) and tests [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
… scaffold [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…sfs_link/aer/amdgpu/kfd [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
… Perfetto export, report, fixture capture, core tests [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…re), replay, Perfetto export, goldens and real-box session fixture [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…rt notice [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…bs, plx status masks), switch counter reasons [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…n blockers, repeats median) [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…helper, port map, CLI, tests, doc section [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…n hip order; session attach [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…captured trees make it unavailable instead of touching live hardware [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…t and GPU name; ports/status columns, hip tag, doc refresh [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…readable); capture stays sysfs-only; tests order- and timing-independent [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…BAR0 reads cause no errors or runtime-PM wake-ups [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
BlivionIaG and others added 28 commits November 12, 2026 17:23
…the live two-lane proof [skip ci]

Fills the measured sections of concurrency-probe-noise-2026-11-12.md from the
five boots on HIP 2-5 (tree 1136bb9 unless named): fixed probe 0 of 20 FAIL
runs on flashnext-mtp2, 0 of 20 on flashnext-mtp0, 0 of 20 with
VLLM_RDNA2_MOE_DETERMINISTIC=1 (solo runs bit-identical in 20 of 20, strict
mode). The known-bad two-lane head twolane-B-adffce5765 fails 5 of 5 runs
(3 or 4 failing checks, both margins far above 1 nat) while its base
twolane-A-3896928d18 passes 5 of 5.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
… the n_dec>=2 repro, stale comment replaced (opengfx1030#165) [skip ci]
…e wvSplitK on gfx10 (opengfx1030#54) [skip ci]

LLMM1 timing on V620 (HIP 0), base 1136bb9 vs PR head 8efea0e, two
ABBA blocks, 14 shapes (fp16/bf16, M 4-8, K 1024-8192): no shape where
both head runs are worse than both base runs; medians -0.5 to -2.0 %.
gfx1030 serving does not reach LLMM1 or wvSplitK (gemv_f16_rdna2 /
rdna_gemm_smallm / rocBLAS); gfx9 code is unchanged (per-function
disassembly diff: only LLGemm1_kernel differs, on gfx1030/1100/1200).
…8 sdot4 config (opengfx1030#89)

The launched kernels are byte-identical on gfx1030 (4 of 4 functions exact
after the trim; the 24 removed instantiations and the probe kernels had no
caller; config 8 was the only selectable id). w4a8_gemm_rdna2 timing on
V620 (HIP 0), base 1136bb9 vs PR head e7ee449, two ABBA blocks, six
prefill shapes: no measurable difference (medians -0.4 to +0.5 %).
…_LEAPQUANT)

Port of LeapQuant (arXiv 2609.38166, vllm#64725) to the RDNA2 GDN path:
the FP32 [V, K] state per value head becomes a 28,768-byte blob (INT8
residual + rank-4 FP16 compensator + FP16 per-token records, window 16).
MTP verify commits accepted records instead of per-token state copies;
align-mode prefix-cache copies commit bias + 1 records in the HIP copy
kernel. Default off; gfx10x only.

decode_kl.py: decode-path top-k KL between served configurations.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
mono_gdn_lq_rdna2 is the fused mono GDN launch (conv + delta rule + gated
norm) with the LeapQuant blob load/append in place of the FP32 state
load/store; gdn_lq_normalize_rdna2 commits the accepted records first. The
flag no longer refuses --monokernel, and the FP32 Triton decode warmup is
skipped under the flag (it never runs there). Shared blob layout moves to
gdn_leapquant_layout.cuh. decode_kl.py accepts a /v1/completions URL.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
All LeapQuant GDN layers normalize with the same state indices and accepted
counts each step, so the first layer of the model now normalizes every
layer's paged blobs in one launch (gdn_lq_normalize_rdna2 gains an optional
layer_offsets [layers] table, grid.y = layer; the group is rebuilt on
bind_kv_cache, never under capture) and the other layers skip. Decode arenas
stay per layer. Replaces 36 launches (~3 us each) and lets the folds of all
layers run concurrently instead of one layer after the other. The record
replay loops unroll by 2 (same arithmetic order).

Per step, 36 layers, one V620, HIP graph (FP32 MonoKernel -> LeapQuant
before -> after): decode n=1 343 -> 574 -> 361 us, verify n=1 ~500 -> 867 ->
582 us, decode n=8 1556 -> 1345 -> 1134 us, verify n=8 3323 -> 2217 ->
1851 us. test_grouped_normalize_matches_per_layer: same bytes as per-layer
normalize, folds and NULL slots included.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…yer group

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…he group

Hybrid models spread their GDN layers over several KV-cache groups, each
with its own block table, so a request sits at different state slots per
group. The grouped normalize (8ba674a) let the first GDN layer of the
model normalize every GDN layer at its own group's slots: the other groups'
layers committed records at foreign slots and never at their own. On
flashnext-mtp2 (6 KV-cache groups) decode fell into repetition loops
(10-21 of 32 outputs at c=8 vs 2-7 for the base) while prefill KL and PPL
stayed inside the floor.

Groups are now built per KV-cache group in bind_kv_cache_to_layers, next to
share_replayssm_ring_trackers (same namespace rule); without the group map
every layer keeps its own launch.

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…ip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
… ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…ffs [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…k pause [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…indow 32, MTP2, F&P)

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…rs and docs [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…=1 and c=8 [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…st [skip ci]

Signed-off-by: BlivionIaG <kev29lt@gmail.com>
…eduler and flashnext-leapquant-mtp2 recipe
Sync the fork with upstream rdna_extras at 3e27433 (2026-11-14), all of it, on the owner's request (2026-11-15):
"Add everything, even the stuff you said doesn't impact our serving path. New models may be added later, so use it
all." The fork's own commits on rdna_extras (the launchers' 90 °C thermal abort and its tests, #2) merge without
conflicts.

The range 750fff8..3e27433 is 119 commits on the same vLLM v0.33 base:
- LeapQuant INT8 GDN recurrent state for Flash-Next (opt-in VLLM_RDNA_GDN_LEAPQUANT, lossy), its admission and
  decoder-first scheduling, and the flashnext-leapquant-mtp2 recipe (512-token blocks, MTP2, MonoKernel) (opengfx1030#168);
- the adaptive mixed-step prefill fit recovering after cold steps (recipes with VLLM_RDNA_DECODE_STALL_MS);
- LLMM1's wave32 reduction fix and wvSplitK refused on gfx10 instead of faulting the queue (opengfx1030#54);
- W4A8 sdot4: only the launched config is built (opengfx1030#89);
- opt-in MoE routing statistics (VLLM_EXPERT_STATS_DIR) and the hetero-MoE placement solver (opengfx1030#158);
- pciescope, the PCIe fabric monitor: passive sysfs/AER/amdgpu/KFD collectors, the PEX88096 switch reader, the
  active probe, a TUI and an MCP server (opengfx1030#157);
- the concurrency probe's noise triage (opengfx1030#167), the GQA decode validation (opengfx1030#165), cleanup follow-ups (opengfx1030#118), and
  upstream's own disaggregation and T10 notes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CnjpEQc6W7ePB6ggy5Tkhe
…one PEX880xx)

pciescope's lease check and port map take a box's layout from its GPU buses, and a box that matches no documented
layout gets NO_LAYOUT: every GPU is refused, as the lease README asks until the table is re-derived. The owner of
jmacpolin/v620-Experimentation wants pciescope on that box (2026-11-15), so it gets its layouts:

- four V620 behind one Broadcom PEX880xx (1000:c010, the chip pciescope's switch reader knows), one lease key
  v620x4 for all four, slot A;
- the buses follow the CPU population: 64 69 6f 72 with two Xeon Gold 6144 (since 2026-10-11), b1 b6 bc bf with
  one, from that repository's builds/v620-cards.tsv (rocm-smi inventories; HIP order = bus order).

Both tables (probe/hw.py LAYOUTS, topology.DOCUMENTED_LAYOUTS) get them, as their own test requires; a new test
covers the key, the lookup by one bus and that a subset is not the box. par1-cs25's layouts are unchanged. The
pciescope suite: 968 passed, 1 skipped (967 + 1 before the change).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CnjpEQc6W7ePB6ggy5Tkhe
@jmacpolin
jmacpolin marked this pull request as ready for review November 15, 2026 17:42
@jmacpolin
jmacpolin merged commit b6bffd1 into rdna_extras Nov 15, 2026
7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants