Migration: `dsv4` → `pocket` / PocketLLM¶

This project was formerly named dsv4, after the first model it ran (DeepSeek-V4). It now supports several unrelated model families, so the project name no longer belongs in the namespace, the build targets, the executable, or the environment variables.

This is a breaking change with no compatibility shim. Old names are gone rather than aliased: a silently-ignored DSV4_* variable would look like a tuning flag that stopped working, which is worse than an obviously-absent one. Translate your commands using the tables below.

What is not renamed¶

DeepSeek-V4 is a model name, not a project name. It is unchanged everywhere it refers to an actual upstream model, checkpoint, architecture, or tensor layout — including the deepseek-v4-flash OpenAI model id, docs/models/deepseek-v4.md, and the historical docs/reports/dsv4_2080ti_report.pdf artifact.

The working directory /mnt/data1/dsv4_inference is also untouched. Renaming a checkout directory is a local decision, and several scripts hardcode that absolute path.

C++ namespace, macros, and build targets¶

Kind	Old	New
Namespace	`dsv4::`	`pocket::`
Test namespace	`dsv4_test`	`pocket_test`
Static library	`dsv4_cpp_core`	`pocket_cpp_core`
Executable	`dsv4_cpp_engine`	`pocketllm_engine`
Build macros	`DSV4_HAVE_TP_COMM`, `DSV4_REQUIRE_NCCL`, `DSV4_LAUNCH_*`, `DSV4_QWEN_MMA_AVAILABLE`	same names with a `POCKET_` prefix

The existing CMake convention is followed: POCKET_* for build-time variables, options, and C++ identifiers; POCKETLLM_* for runtime environment variables. The Python extension module was already pocketllm_cpp and does not change.

pocket_cpp_core is a static archive with no install() rule and no SOVERSION, so there is no ABI or soname consequence for the rename.

Files renamed¶

Old	New
`cpp_engine/engine/dsv4_engine.cpp`	`cpp_engine/engine/deepseek_v4_engine.cpp`
`cpp_engine/include/dsv4_engine.hpp`	`cpp_engine/include/deepseek_v4_engine.hpp`
`src/encoding/dsv4.py`	`src/encoding/deepseek_v4.py`
`tests/test_encoding_dsv4.py`	`tests/test_encoding_deepseek_v4.py`

These four keep the model name because that is what they implement; only the abbreviation is expanded. The C++ class Dsv4Engine is now DeepSeekV4Engine, and the Python import src.encoding.dsv4 is now src.encoding.deepseek_v4.

Runtime and wire identifiers¶

Kind	Old	New
Shared-memory segment	`dsv4_cpu_moe_server`	`pocketllm_cpu_moe_server`
In-process shm prefix	`dsv4_cpu_moe_inproc_`	`pocketllm_cpu_moe_inproc_`
MoE weight shm prefix	`dsv4_moe_`	`pocketllm_moe_`
MoE shm header magic	`DSV4MOE1`	`PKTLMOE1`
GGUF capture tag	`dsv4_runtime_capture`	`pocketllm_runtime_capture`
Attention kind label	`dsv4_mla_compressed_sparse`	`deepseek_v4_mla_compressed_sparse`
CUDA device function	`dsv4_dp4a_i8`	`pocket_dp4a_i8`

The shared-memory names are still overridable through DEEPSEEK_CPU_MOE_SERVER_SHM, which is unchanged. The header magic is the same width (8 bytes), so a stale segment left by an old process is rejected rather than misread. The GGUF capture tag is only ever written into metadata, and the attention-kind label is only ever printed — neither is matched anywhere, so neither affects dispatch or the ability to read existing files.

Scratch paths, lock files, log names, and NCCL rendezvous files under /tmp moved from dsv4_* to pocketllm_*. These are per-run temporaries; nothing persists across the rename.

Environment variables¶

Every DSV4_* variable is now POCKETLLM_* with the rest of the name unchanged, so the mechanical translation is a prefix substitution:

# before
DSV4_QWEN_GQA_OPTIMIZED=1 DSV4_CPP_PREFILL_BATCHED_ATTN=1 ./dsv4_cpp_engine ...
# after
POCKETLLM_QWEN_GQA_OPTIMIZED=1 POCKETLLM_CPP_PREFILL_BATCHED_ATTN=1 ./pocketllm_engine ...

QWEN_*, GLM_*, DEEPSEEK_*, and TP_* variables are not affected.

To rewrite a saved script in place:

sed -i 's/\bDSV4_/POCKETLLM_/g' your_script.sh

The complete list of the 161 renamed variables follows.

Old	New
`DSV4_BENCH_MODE`	`POCKETLLM_BENCH_MODE`
`DSV4_CKPT_DIR`	`POCKETLLM_CKPT_DIR`
`DSV4_CPP_BATCHED_VERIFY`	`POCKETLLM_CPP_BATCHED_VERIFY`
`DSV4_CPP_DEBUG_FORWARD`	`POCKETLLM_CPP_DEBUG_FORWARD`
`DSV4_CPP_DECODE_PROFILE`	`POCKETLLM_CPP_DECODE_PROFILE`
`DSV4_CPP_DECODE_SPARSE_ARENA`	`POCKETLLM_CPP_DECODE_SPARSE_ARENA`
`DSV4_CPP_DECODE_WO_A_INT8`	`POCKETLLM_CPP_DECODE_WO_A_INT8`
`DSV4_CPP_DENSE_ARENA_MAX_LAYERS`	`POCKETLLM_CPP_DENSE_ARENA_MAX_LAYERS`
`DSV4_CPP_EOS_TOKEN_ID`	`POCKETLLM_CPP_EOS_TOKEN_ID`
`DSV4_CPP_GPU_COMPRESSOR`	`POCKETLLM_CPP_GPU_COMPRESSOR`
`DSV4_CPP_MOE_DETERMINISTIC_REDUCE`	`POCKETLLM_CPP_MOE_DETERMINISTIC_REDUCE`
`DSV4_CPP_MOE_FORCE_PADDED`	`POCKETLLM_CPP_MOE_FORCE_PADDED`
`DSV4_CPP_MOE_NSPLIT`	`POCKETLLM_CPP_MOE_NSPLIT`
`DSV4_CPP_MOE_WMMA_WARPS`	`POCKETLLM_CPP_MOE_WMMA_WARPS`
`DSV4_CPP_NCCL_ID_WAIT_ATTEMPTS`	`POCKETLLM_CPP_NCCL_ID_WAIT_ATTEMPTS`
`DSV4_CPP_PREFILL_BATCHED_ATTN`	`POCKETLLM_CPP_PREFILL_BATCHED_ATTN`
`DSV4_CPP_PREFILL_CHUNK`	`POCKETLLM_CPP_PREFILL_CHUNK`
`DSV4_CPP_PREFILL_CHUNK_TOKENS`	`POCKETLLM_CPP_PREFILL_CHUNK_TOKENS`
`DSV4_CPP_PREFILL_GPU_HASH_GATE`	`POCKETLLM_CPP_PREFILL_GPU_HASH_GATE`
`DSV4_CPP_PREFILL_MOE_COPY_STREAM`	`POCKETLLM_CPP_PREFILL_MOE_COPY_STREAM`
`DSV4_CPP_PREFILL_MOE_PREFETCH`	`POCKETLLM_CPP_PREFILL_MOE_PREFETCH`
`DSV4_CPP_PREFILL_PRESTAGE_EXPERTS`	`POCKETLLM_CPP_PREFILL_PRESTAGE_EXPERTS`
`DSV4_CPP_PREFILL_ROUTE_STATS`	`POCKETLLM_CPP_PREFILL_ROUTE_STATS`
`DSV4_CPP_PREFILL_SHARED_OVERLAP`	`POCKETLLM_CPP_PREFILL_SHARED_OVERLAP`
`DSV4_CPP_PREFILL_WARMUP_PASSES`	`POCKETLLM_CPP_PREFILL_WARMUP_PASSES`
`DSV4_CPP_PREPARE_FP4_HOST`	`POCKETLLM_CPP_PREPARE_FP4_HOST`
`DSV4_CPP_PREPARE_PROGRESS`	`POCKETLLM_CPP_PREPARE_PROGRESS`
`DSV4_CPP_PROFILE_ATTN`	`POCKETLLM_CPP_PROFILE_ATTN`
`DSV4_CPP_PROFILE_FORWARD`	`POCKETLLM_CPP_PROFILE_FORWARD`
`DSV4_CPP_PROFILE_MOE_HOST`	`POCKETLLM_CPP_PROFILE_MOE_HOST`
`DSV4_CPP_PROFILE_MOE_PREFILL`	`POCKETLLM_CPP_PROFILE_MOE_PREFILL`
`DSV4_CPP_PROFILE_MOE_STAGE`	`POCKETLLM_CPP_PROFILE_MOE_STAGE`
`DSV4_CPP_PROFILE_REDUCE_DETAIL`	`POCKETLLM_CPP_PROFILE_REDUCE_DETAIL`
`DSV4_CPP_TOPK_DIAG`	`POCKETLLM_CPP_TOPK_DIAG`
`DSV4_DFLASH2_ADAPTIVE_WIDTH`	`POCKETLLM_DFLASH2_ADAPTIVE_WIDTH`
`DSV4_DFLASH2_CUBLAS_FP32`	`POCKETLLM_DFLASH2_CUBLAS_FP32`
`DSV4_DFLASH2_DOWNLOAD_CANDIDATES`	`POCKETLLM_DFLASH2_DOWNLOAD_CANDIDATES`
`DSV4_DFLASH2_DRAFT_WIDTH`	`POCKETLLM_DFLASH2_DRAFT_WIDTH`
`DSV4_DFLASH2_FUSED_SWIGLU`	`POCKETLLM_DFLASH2_FUSED_SWIGLU`
`DSV4_DFLASH2_GROUPED_ATTN`	`POCKETLLM_DFLASH2_GROUPED_ATTN`
`DSV4_DFLASH2_PROFILE`	`POCKETLLM_DFLASH2_PROFILE`
`DSV4_DFLASH2_REFINE_PASSES`	`POCKETLLM_DFLASH2_REFINE_PASSES`
`DSV4_DFLASH2_SHARD`	`POCKETLLM_DFLASH2_SHARD`
`DSV4_DFLASH2_SMALL_BATCH_GEMM`	`POCKETLLM_DFLASH2_SMALL_BATCH_GEMM`
`DSV4_DFLASH2_SPLIT_TOPK`	`POCKETLLM_DFLASH2_SPLIT_TOPK`
`DSV4_DFLASH2_VITERBI_SELECTOR`	`POCKETLLM_DFLASH2_VITERBI_SELECTOR`
`DSV4_DSPARK_LM_HEAD_CUBLAS`	`POCKETLLM_DSPARK_LM_HEAD_CUBLAS`
`DSV4_GGUF_CHUNKED_PREFILL`	`POCKETLLM_GGUF_CHUNKED_PREFILL`
`DSV4_GGUF_DEBUG_LOGIT_TOKENS`	`POCKETLLM_GGUF_DEBUG_LOGIT_TOKENS`
`DSV4_GGUF_DECODE_ONLY_ATTENTION_ONLY`	`POCKETLLM_GGUF_DECODE_ONLY_ATTENTION_ONLY`
`DSV4_GGUF_DECODE_ONLY_CONTEXT`	`POCKETLLM_GGUF_DECODE_ONLY_CONTEXT`
`DSV4_GGUF_DECODE_ONLY_WARMUP`	`POCKETLLM_GGUF_DECODE_ONLY_WARMUP`
`DSV4_GGUF_DEVICE_ROUTE_SLOTS`	`POCKETLLM_GGUF_DEVICE_ROUTE_SLOTS`
`DSV4_GGUF_FLASHMEMORY_CKPT`	`POCKETLLM_GGUF_FLASHMEMORY_CKPT`
`DSV4_GGUF_FLASHMEMORY_ENSEMBLE`	`POCKETLLM_GGUF_FLASHMEMORY_ENSEMBLE`
`DSV4_GGUF_FLASHMEMORY_LOAD_DEVICE`	`POCKETLLM_GGUF_FLASHMEMORY_LOAD_DEVICE`
`DSV4_GGUF_FLASHMEMORY_METADATA_ONLY`	`POCKETLLM_GGUF_FLASHMEMORY_METADATA_ONLY`
`DSV4_GGUF_FLASHMEMORY_MOCK_SCORER_SMOKE`	`POCKETLLM_GGUF_FLASHMEMORY_MOCK_SCORER_SMOKE`
`DSV4_GGUF_FLASHMEMORY_PLUGIN`	`POCKETLLM_GGUF_FLASHMEMORY_PLUGIN`
`DSV4_GGUF_FLASHMEMORY_RUNTIME_SCORING`	`POCKETLLM_GGUF_FLASHMEMORY_RUNTIME_SCORING`
`DSV4_GGUF_FLASHMEMORY_SRC_LAYER`	`POCKETLLM_GGUF_FLASHMEMORY_SRC_LAYER`
`DSV4_GGUF_HASH_TABLE_RESIDENT`	`POCKETLLM_GGUF_HASH_TABLE_RESIDENT`
`DSV4_GGUF_HOST_LOGITS`	`POCKETLLM_GGUF_HOST_LOGITS`
`DSV4_GGUF_INDEXED_ATTN`	`POCKETLLM_GGUF_INDEXED_ATTN`
`DSV4_GGUF_INDEXED_ATTN_AUTO_THRESHOLD`	`POCKETLLM_GGUF_INDEXED_ATTN_AUTO_THRESHOLD`
`DSV4_GGUF_INDEXED_ATTN_WINDOW`	`POCKETLLM_GGUF_INDEXED_ATTN_WINDOW`
`DSV4_GGUF_IQ1_FUSED_W13_SWIGLU`	`POCKETLLM_GGUF_IQ1_FUSED_W13_SWIGLU`
`DSV4_GGUF_IQ1_W2_REDUCE`	`POCKETLLM_GGUF_IQ1_W2_REDUCE`
`DSV4_GGUF_KV_SWAP`	`POCKETLLM_GGUF_KV_SWAP`
`DSV4_GGUF_KV_SWAP_GPU_CHUNKS`	`POCKETLLM_GGUF_KV_SWAP_GPU_CHUNKS`
`DSV4_GGUF_KV_SWAP_PINNED_MIB`	`POCKETLLM_GGUF_KV_SWAP_PINNED_MIB`
`DSV4_GGUF_KV_SWAP_SYNC`	`POCKETLLM_GGUF_KV_SWAP_SYNC`
`DSV4_GGUF_KV_SWAP_TEST_ALLOW_TOPK_REDUCTION`	`POCKETLLM_GGUF_KV_SWAP_TEST_ALLOW_TOPK_REDUCTION`
`DSV4_GGUF_KV_SWAP_VALIDATE`	`POCKETLLM_GGUF_KV_SWAP_VALIDATE`
`DSV4_GGUF_LOAD_ONLY`	`POCKETLLM_GGUF_LOAD_ONLY`
`DSV4_GGUF_MEM_PROFILE`	`POCKETLLM_GGUF_MEM_PROFILE`
`DSV4_GGUF_MIN_FREE_MIB`	`POCKETLLM_GGUF_MIN_FREE_MIB`
`DSV4_GGUF_MOE_COPY_STREAM`	`POCKETLLM_GGUF_MOE_COPY_STREAM`
`DSV4_GGUF_PINNED_STAGE`	`POCKETLLM_GGUF_PINNED_STAGE`
`DSV4_GGUF_PINNED_STAGE_CAP_MIB`	`POCKETLLM_GGUF_PINNED_STAGE_CAP_MIB`
`DSV4_GGUF_PINNED_STAGE_SLOTS`	`POCKETLLM_GGUF_PINNED_STAGE_SLOTS`
`DSV4_GGUF_PREFILL_ATTN_MAX_INDEX_ELEMS`	`POCKETLLM_GGUF_PREFILL_ATTN_MAX_INDEX_ELEMS`
`DSV4_GGUF_PREFILL_ATTN_WINDOW`	`POCKETLLM_GGUF_PREFILL_ATTN_WINDOW`
`DSV4_GGUF_PREFILL_CHUNK_TOKENS`	`POCKETLLM_GGUF_PREFILL_CHUNK_TOKENS`
`DSV4_GGUF_PREFILL_HEADPAIR_ATTN`	`POCKETLLM_GGUF_PREFILL_HEADPAIR_ATTN`
`DSV4_GGUF_PREFILL_HEADPAIR_SERIAL_ATTN`	`POCKETLLM_GGUF_PREFILL_HEADPAIR_SERIAL_ATTN`
`DSV4_GGUF_PREFILL_INDEXED_ATTN`	`POCKETLLM_GGUF_PREFILL_INDEXED_ATTN`
`DSV4_GGUF_PREFILL_INDEXED_ATTN_MAX_TOPK`	`POCKETLLM_GGUF_PREFILL_INDEXED_ATTN_MAX_TOPK`
`DSV4_GGUF_PREFILL_REFINE_LAST`	`POCKETLLM_GGUF_PREFILL_REFINE_LAST`
`DSV4_GGUF_PREFILL_REFINE_LAST_TOKENS`	`POCKETLLM_GGUF_PREFILL_REFINE_LAST_TOKENS`
`DSV4_GGUF_PREFILL_SHARED_OVERLAP`	`POCKETLLM_GGUF_PREFILL_SHARED_OVERLAP`
`DSV4_GGUF_PROFILE`	`POCKETLLM_GGUF_PROFILE`
`DSV4_GGUF_REDUCE_FP32`	`POCKETLLM_GGUF_REDUCE_FP32`
`DSV4_GGUF_RESIDENT_ROUTED_EXPERTS`	`POCKETLLM_GGUF_RESIDENT_ROUTED_EXPERTS`
`DSV4_GGUF_RESIDENT_ROUTED_LAYERS`	`POCKETLLM_GGUF_RESIDENT_ROUTED_LAYERS`
`DSV4_GGUF_SPARSE_ATTN_THRESHOLD`	`POCKETLLM_GGUF_SPARSE_ATTN_THRESHOLD`
`DSV4_GGUF_SPARSE_COMPRESSOR`	`POCKETLLM_GGUF_SPARSE_COMPRESSOR`
`DSV4_GGUF_SPARSE_INDEXER`	`POCKETLLM_GGUF_SPARSE_INDEXER`
`DSV4_GGUF_SPARSE_WINDOW`	`POCKETLLM_GGUF_SPARSE_WINDOW`
`DSV4_GGUF_VERBOSE`	`POCKETLLM_GGUF_VERBOSE`
`DSV4_IQ1_GROUPED_GEMM`	`POCKETLLM_IQ1_GROUPED_GEMM`
`DSV4_IQ1_GROUPED_GEMM_NSPLIT`	`POCKETLLM_IQ1_GROUPED_GEMM_NSPLIT`
`DSV4_IQ1_GROUPED_GEMM_W13`	`POCKETLLM_IQ1_GROUPED_GEMM_W13`
`DSV4_IQ1_GROUPED_GEMM_W2`	`POCKETLLM_IQ1_GROUPED_GEMM_W2`
`DSV4_IQ1_GROUPED_PROFILE`	`POCKETLLM_IQ1_GROUPED_PROFILE`
`DSV4_IQ1_GROUPED_ROUTE_MAJOR`	`POCKETLLM_IQ1_GROUPED_ROUTE_MAJOR`
`DSV4_IQ1_GROUPED_ROUTE_TILE4`	`POCKETLLM_IQ1_GROUPED_ROUTE_TILE4`
`DSV4_IQ1_GROUPED_W2_Q8`	`POCKETLLM_IQ1_GROUPED_W2_Q8`
`DSV4_Q2_PREFILL_W13_EXPERT_TILE`	`POCKETLLM_Q2_PREFILL_W13_EXPERT_TILE`
`DSV4_Q2_PREFILL_W2_EXPERT_TILE`	`POCKETLLM_Q2_PREFILL_W2_EXPERT_TILE`
`DSV4_Q2_PREFILL_W2_ROUTE_TILE`	`POCKETLLM_Q2_PREFILL_W2_ROUTE_TILE`
`DSV4_Q8_ROWS_BATCH4`	`POCKETLLM_Q8_ROWS_BATCH4`
`DSV4_QWEN4_ATTN_PHASE_PROFILE`	`POCKETLLM_QWEN4_ATTN_PHASE_PROFILE`
`DSV4_QWEN4_CPU_AFFINITY`	`POCKETLLM_QWEN4_CPU_AFFINITY`
`DSV4_QWEN4_DENSE_FP16`	`POCKETLLM_QWEN4_DENSE_FP16`
`DSV4_QWEN4_DENSE_FP16_FP32_OUTPUT`	`POCKETLLM_QWEN4_DENSE_FP16_FP32_OUTPUT`
`DSV4_QWEN4_DENSE_FP16_MIN_ROWS`	`POCKETLLM_QWEN4_DENSE_FP16_MIN_ROWS`
`DSV4_QWEN4_GDN_CUDA`	`POCKETLLM_QWEN4_GDN_CUDA`
`DSV4_QWEN4_GDN_GROUPS_PER_CTA`	`POCKETLLM_QWEN4_GDN_GROUPS_PER_CTA`
`DSV4_QWEN4_HC_CUDA`	`POCKETLLM_QWEN4_HC_CUDA`
`DSV4_QWEN4_HC_CUDA_MIN_ROWS`	`POCKETLLM_QWEN4_HC_CUDA_MIN_ROWS`
`DSV4_QWEN4_HC_FP16`	`POCKETLLM_QWEN4_HC_FP16`
`DSV4_QWEN4_HC_FP16_FP32_OUTPUT`	`POCKETLLM_QWEN4_HC_FP16_FP32_OUTPUT`
`DSV4_QWEN4_HC_FP16_INJECT`	`POCKETLLM_QWEN4_HC_FP16_INJECT`
`DSV4_QWEN4_HC_FP16_MIN_ROWS`	`POCKETLLM_QWEN4_HC_FP16_MIN_ROWS`
`DSV4_QWEN4_HC_FP16_MIX_DOWN`	`POCKETLLM_QWEN4_HC_FP16_MIX_DOWN`
`DSV4_QWEN4_HC_FP16_MIX_UP`	`POCKETLLM_QWEN4_HC_FP16_MIX_UP`
`DSV4_QWEN4_MOE_FP16_COMPUTE`	`POCKETLLM_QWEN4_MOE_FP16_COMPUTE`
`DSV4_QWEN4_MOE_FP16_MIN_TOKENS`	`POCKETLLM_QWEN4_MOE_FP16_MIN_TOKENS`
`DSV4_QWEN4_MOE_GROUPED`	`POCKETLLM_QWEN4_MOE_GROUPED`
`DSV4_QWEN4_MOE_GROUPED_MAX_PADDING`	`POCKETLLM_QWEN4_MOE_GROUPED_MAX_PADDING`
`DSV4_QWEN4_MOE_GROUPED_MIN_FREE_GIB`	`POCKETLLM_QWEN4_MOE_GROUPED_MIN_FREE_GIB`
`DSV4_QWEN4_MOE_GROUPED_MIN_TOKENS`	`POCKETLLM_QWEN4_MOE_GROUPED_MIN_TOKENS`
`DSV4_QWEN4_MOE_PHASE_PROFILE`	`POCKETLLM_QWEN4_MOE_PHASE_PROFILE`
`DSV4_QWEN4_MOE_PREFETCH`	`POCKETLLM_QWEN4_MOE_PREFETCH`
`DSV4_QWEN4_PROFILE_TREE`	`POCKETLLM_QWEN4_PROFILE_TREE`
`DSV4_QWEN4_QSA_CUDA`	`POCKETLLM_QWEN4_QSA_CUDA`
`DSV4_QWEN4_QSA_FALLBACK_ROWS`	`POCKETLLM_QWEN4_QSA_FALLBACK_ROWS`
`DSV4_QWEN4_QSA_MMA`	`POCKETLLM_QWEN4_QSA_MMA`
`DSV4_QWEN_DECODE_GROUP_HEADS`	`POCKETLLM_QWEN_DECODE_GROUP_HEADS`
`DSV4_QWEN_DECODE_MAX_SPLITS`	`POCKETLLM_QWEN_DECODE_MAX_SPLITS`
`DSV4_QWEN_DECODE_MMA`	`POCKETLLM_QWEN_DECODE_MMA`
`DSV4_QWEN_DECODE_MMA_TARGET_SPLITS`	`POCKETLLM_QWEN_DECODE_MMA_TARGET_SPLITS`
`DSV4_QWEN_DECODE_SPLIT_POSITIONS`	`POCKETLLM_QWEN_DECODE_SPLIT_POSITIONS`
`DSV4_QWEN_DECODE_TARGET_SPLITS`	`POCKETLLM_QWEN_DECODE_TARGET_SPLITS`
`DSV4_QWEN_GQA_FLASH_TILE`	`POCKETLLM_QWEN_GQA_FLASH_TILE`
`DSV4_QWEN_GQA_HEADS_PER_GROUP`	`POCKETLLM_QWEN_GQA_HEADS_PER_GROUP`
`DSV4_QWEN_GQA_LONG_TILE`	`POCKETLLM_QWEN_GQA_LONG_TILE`
`DSV4_QWEN_GQA_MMA_OCC`	`POCKETLLM_QWEN_GQA_MMA_OCC`
`DSV4_QWEN_GQA_MMA_TILE`	`POCKETLLM_QWEN_GQA_MMA_TILE`
`DSV4_QWEN_GQA_OPTIMIZED`	`POCKETLLM_QWEN_GQA_OPTIMIZED`
`DSV4_QWEN_GQA_POS_TILE`	`POCKETLLM_QWEN_GQA_POS_TILE`
`DSV4_QWEN_GQA_QUERY_ROWS`	`POCKETLLM_QWEN_GQA_QUERY_ROWS`
`DSV4_QWEN_GQA_WARP_COMBO`	`POCKETLLM_QWEN_GQA_WARP_COMBO`
`DSV4_QWEN_KV_CACHE_DTYPE`	`POCKETLLM_QWEN_KV_CACHE_DTYPE`
`DSV4_QWEN_NVFP4`	`POCKETLLM_QWEN_NVFP4`
`DSV4_QWEN_NVFP4_FUSED_SWIGLU`	`POCKETLLM_QWEN_NVFP4_FUSED_SWIGLU`
`DSV4_QWEN_NVFP4_SHARED_Q8_SWIGLU`	`POCKETLLM_QWEN_NVFP4_SHARED_Q8_SWIGLU`
`DSV4_QWEN_NVFP4_WIDE_N64`	`POCKETLLM_QWEN_NVFP4_WIDE_N64`
`DSV4_QWEN_NVFP4_WIDE_N64_MIN_ROWS`	`POCKETLLM_QWEN_NVFP4_WIDE_N64_MIN_ROWS`
`DSV4_TMP_DIR`	`POCKETLLM_TMP_DIR`

Migration: dsv4 → pocket / PocketLLM¶